Ultrasonographically detected FLLs in dogs are a diagnostic dilemma in veterinary medical imaging because of the difficulty in separating lesions of malignant or benign diseases. Ultrasonographic criteria for making this distinction have not been adequately investigated in veterinary medicine. It is widely accepted that the ultrasonographic characteristics of FLLs cannot assist in identifying the specific cause of most lesions1–4 and that invasive sampling procedures such as collection of fine-needle aspirates or biopsy specimens are required for diagnosis.5,6 This belief arises in large part from the wide variety of appearances of neoplasms that affect the liver. For example, the appearance of HCCs can be in 3 forms: mass form (single large mass), nodular form (multiple nodules in ≥ 1 lobe), and diffuse or infiltrative form.7 Bile duct carcinomas can also have the 3 forms of appearance seen with HCCs.8 Lymphosarcoma may have an ultrasonographically diffuse form (which causes decreased liver echogenicity), a poorly defined hypoechoic to anechoic form, and a nodular pattern consisting of multiple round hyperechoic foci surrounded by hypoechoic areas.9
Ultrasonographic features of lesions attributable to malignant and benign diseases of the liver may overlap. Nodular hyperplasia can have an ultrasonographic appearance that can easily be mistaken for that of neoplasia.10 Target lesions, which consist of a focal lesion with a hyperechoic center surrounded by a hypoechoic rim, were originally considered to be strongly associated with malignancy, but they have also been found to be attributable to benign conditions.11
The assertion that the ultrasonographic features of FLLs are of little value for use in determining their cause can be traced to early veterinary studies12–16 on the ultrasonographic appearance of liver cancer, studies17,18 in humans in which the various characteristics of focal liver diseases are described, and veterinary studies10,11,19–25 in which investigators describe a wide variety of diseases that may cause ultrasonographically detectable FLLs. However, in 1 early study16 of the ultrasonographic features of hepatic neoplasms, the investigators detected patterns in the appearance of some lesions for malignant conditions. Focal lesions were found to be more likely to be HCCs, whereas diffuse changes were more likely to represent lymphosarcoma.16 In a more recent study,26 it was suggested that certain ultrasonographic features may be associated with the cytologic diagnosis of liver lesions. Therefore, although the ultrasonographic features of FLLs may not allow clinicians to provide a specific diagnosis, they may provide information about the more general question of whether an FLL is attributable to a malignant or benign condition.
In human medicine, the ultrasonographic characteristics of liver lesions, although acknowledged as nonspecific, are combined with other patient data to guide clinical decision making.27,28 In dogs with liver disease, these other patient data (consisting of clinical signs, history, and laboratory findings) are generally nonspecific.29 There may be CNS signs, such as seizures or behavior changes. Laboratory findings may or may not reflect liver disease or occasionally may reflect a particular tumor type.30 There are also reports31,32 of laboratory abnormalities in patients with neoplastic disease, but use of these abnormalities alone is insufficiently accurate for clinical use.
It is clear that clinical characteristics or ultrasonographic findings are not specific enough for use in determining the cause of ultrasonographically detected FLLs. However, a combination of clinical, laboratory, and imaging findings may allow prediction of whether lesions are more likely to be attributable to benign or malignant conditions. Determining whether a patient is likely to have benign or malignant disease could aid clinicians in deciding whether invasive procedures for collection of samples are warranted. In addition, knowledge of the relative contribution of various types of diagnostic data to a confirmed diagnosis would allow relative weighting of these data.
The purpose of the study reported here was to determine the relative predictive performance of multivariable logistic regression models for combinations of patient data (defined as signalment, owner-reported medical history, physical examination findings, laboratory data, and thoracic radiographic findings) and ultrasonographic data to predict whether ultrasonographically detected FLLs in dogs were associated with malignant or benign conditions (as determined on the basis of histologic results). We also intended to identify those predictor variables that had the most relevance to the predictive ability of the statistical models, which could translate to the relevancy of these data with regard to the differentiation of liver lesions attributable to malignant versus benign diseases.
Materials and Methods
Sample—A retrospective cross-sectional study was conducted with information for dogs undergoing abdominal ultrasonographic examination from February 2005 through October 2008 at the University of Minnesota Veterinary Medical Center. Dogs were selected for inclusion in the study if they met the following criteria: adequate representative archived static ultrasonographic images of the liver with confirmation of FLL and reports for histologic examination of samples of the liver obtained via needle core biopsy, during surgery, or during necropsy within 14 days before or after the ultrasonographic examination. Only results of histologic examination conducted at the University of Minnesota Veterinary Medical Center laboratory or a national diagnostic laboratorya were used for the study.
A total of 4,664 accessions for abdominal ultrasonographic examination during the study period were identified. Reports of ultrasonographic examinations were reviewed by 1 author (TM) for reference to FLLs. Such references included, but were not limited to, mention of the presence of nodules, masses, irregular margins, or a nodular appearance. Of the 4,664 accessions, 1,122 were associated with reports that mentioned FLLs. Of these, 247 met the inclusion criteria.
Candidate predictor variables extracted from the medical records included signalment data, owner-reported medical history, physical examination findings, laboratory test results, thoracic radiographic findings, and abdominal ultrasonographic findings. Signalment data consisted of age and sex. Owner-reported medical history consisted of duration of clinical signs; vomiting, diarrhea, or a decrease in appetite within 10 days before the abdominal ultrasonographic examination; or administration of medications that potentially could have affected the liver. For the purposes of the study, medications considered as having the potential to affect the liver were limited to glucocorticoids, antiepileptic drugs, and carprofen. Physical examination results recorded for the study consisted of peripheral lymphadenopathy, palpable abdominal organomegaly, ascites, jaundice, or fever (rectal temperature > 38.9°C). Except for the duration of clinical signs, historical and physical examination findings were considered absent if not recorded in the medical records.
Thoracic radiographs were evaluated by 2 board-certified veterinary radiologists (TM and DAF) who were unaware of the nonimaging data for the dogs. Thoracic radiographs and abdominal ultrasonographic images of each dog were evaluated by the 2 radiologists, with the thoracic radiographs evaluated first, followed by evaluation of the abdominal ultrasonographic images. Ultrasonographic images were not made available during evaluation of the thoracic radiographs. Survey thoracic radiographs were evaluated only if obtained within 10 days before or after the ultrasonographic examination. A 10-day interval was chosen because of the chronic nature of most diseases that cause FLLs, which were considered unlikely to lead to dramatic radiographic changes over such a short interval. The radiographs were evaluated concurrently by both radiologists, who reached a consensus regarding the presence or absence of pulmonary nodules, abnormal liver (irregular margins, focal enlargement, or diffuse enlargement), or decreased peritoneal detail.
Laboratory test results extracted from the medical records consisted of the presence or absence of hypoalbuminemia, hyperbilirubinemia, increased ALP or ALT activity, and increased coagulation time (increase in prothrombin time or partial thromboplastin time). These results were included in the study if the tests had been performed within 10 days before or after the ultrasonographic examination and if the tests had been performed at the University of Minnesota Veterinary Medical Center. A 10-day interval was chosen because the disease processes of interest were generally chronic in nature. A complete set of data (ie, signalment, medical history, results of physical examination, results of thoracic radiography, and laboratory test results) were not required for inclusion in the study.
The ultrasonographic features evaluated as potential predictors of histologic findings indicative of malignant versus benign conditions for the liver consisted of the size of FLLs, number of FLLs (1 lesion, 2 to 5 lesions, > 5 discrete lesions, or complete effacement of the liver by FLLs), variation in size of FLLs (a solitary lesion or a difference of < 50% between the diameters of the smallest and largest FLLs, or a difference of > 50% between the diameters of the smallest and largest FLLs), FLL echogenicity relative to liver parenchymal echogenicity (isoechoic, hyperechoic, hypoechoic, heteroechoic, or hypoechoic with distal acoustic enhancement), and presence or absence of a heterogeneous (coarse) liver parenchyma, a hyperechoic liver (greater echogenicity than that of the spleen), abdominal lymphadenopathy, or an abdominal mass (any mass arising from an abdominal organ other than the liver). The archived static abdominal ultrasonographic images were evaluated by the same investigators who evaluated the thoracic radiographs (TM and DAF), and the presence or absence of imaging findings was established by consensus.
Causes of the FLLs were established on the basis of pathology reports from the University of Minnesota Veterinary Medical Center or a national laboratory.a The classification of benign versus malignant disease was made on the basis of guidelines for the classification of liver disease in dogs and cats.33 Lesions were classified as malignant disease if the microscopic findings indicated round cell neoplasia (lymphoma, mast cell tumor, or histiocytic tumor), primary liver tumors such as HCC (regardless of grade) or cholangiocellular carcinomas, or metastatic neoplasms. Lesions were classified as benign disease if the pathology report indicated nodular regeneration, cirrhosis, inflammatory diseases, benign tumors such as hepatic adenoma, or a histologically normal liver. Those dogs in which a well-differentiated HCC was listed as a differential diagnosis but not specifically diagnosed as an HCC were classified as having benign disease. Because it was not possible to establish the exact liver lesions that had been examined microscopically in the present study, samples for histologic examination were viewed as representative samples of the liver.
Statistical analysis—Because thoracic radiographs and laboratory tests were not inclusion criteria for this study, some dogs had missing data for these categories. To determine whether the missing patient data were related to outcome, variables for missing laboratory tests (serum biochemical analysis, a CBC, or a coagulation panel) and missing thoracic radiographs were created. Data were categorized as missing if an entire diagnostic test group was missing. If individual variables within the diagnostic test group were missing (eg, if a serum biochemical analysis did not include a particular liver enzyme test) but the test was available, that dog's data was classified as not missing for that diagnostic test. The relationship between missing data and outcome was assessed for the entire data set by means of a Fisher exact test.
The data set was randomly divided into development and validation data sets. A random number generator in a publicly available statistical programb was used to assign approximately 70% of the dogs to the development data set, with the remaining 30% comprising the validation data set. For missing data, multiple imputations of the missing data were created by use of chained equations. Briefly, multiple complete versions of the data set were imputed from the incomplete data set by selecting plausible values for the missing data from specifically modeled distributions for the missing data.34 Statistical analyses were performed on the multiple-imputed data sets, and results from each imputed data set were pooled for the final results. Multiple imputations were performed separately for the development and validation data sets. Signalment data, owner-reported medical history, physical examination findings, laboratory test results, thoracic radiographic findings, and outcome of histologic examination (benign vs malignant) were used as predictors for the missing data. Twenty-five imputations were performed for both the development and validation data sets, which was a conservatively large number of imputations on the basis of the percentage of missing data for some of the variables.35
Three multivariable logistic regression models were fit to the multiple-imputed development data set by use of predictor variables selected from various categories of candidate predictor variables via a backward stepwise selection process. The first model (full model) incorporated selected candidate predictor variables from all patient data assessed in the study (signalment, owner-reported medical history, physical examination findings, laboratory test results, thoracic radiographic findings, and abdominal ultrasonographic findings). The second model (ultrasonographic model) incorporated selected candidate predictors from only the abdominal ultrasonographic findings. The third model (clinical model) incorporated selected candidate predictors from all categories of patient information, with the exclusion of the abdominal ultrasonographic findings. A backward stepwise selection process conducted with the Akaike information criterion was used to select variables with the strongest association with malignant versus benign results for histologic examination. The Akaike information criterion–based selection process was performed on each of the 25 imputed development data sets. A majority method was used to obtain the final models (full, ultrasonographic, and clinical) in which all variables that were selected via the Akaike information criterion in > 50% of the imputed data sets (ie, ≥ 13/25 impuations) were included.35 Pooled regression coefficients were obtained for all 3 models fit to the development data set by use of Rubin rules36 as implemented in a publically available statistical package.34 Performance measures against the development and validation data sets were obtained by calculating the mean of the individual measures for each model fit to the imputed data sets.37 The c-index (equivalent to the area under a receiver operating characteristic curve) and the Nagelkerke R2 were obtained from model fits of both the development and validation data sets to assess model performance. The 3 models were also used to obtain predictions from the validation data set. In addition to the c-index and Nagelkerke R2, calibration was assessed from calibration plots. The mean intercepts and slopes from the calibration plots against the 25 imputed data sets were obtained. Perfect calibration would result in an intercept of 0 and a slope of 1.35
Study data were entered into a commercially available spreadsheet program.c All statistical analyses were performed by use of a publicly available statistical software program.b Values of P < 0.05 were considered significant.
Results
The 247 dogs included in the study ranged from 1.5 to 16.5 years of age (mean ± SD, 10.8 ± 2.4 years). Dogs represented 62 breeds, including 48 Labrador Retrievers, 21 Golden Retrievers, 14 Beagles, 12 German Shepherd Dogs, 11 Cocker Spaniels, 11 Bichon Frises, 8 Shetland Sheepdogs, and 12 mixed-breed dogs. The remaining 55 breeds were represented by ≤ 6 dogs/breed. There were 107 neutered males, 124 spayed females, 12 sexually intact males, and 4 sexually intact females.
Histologic confirmation was obtained by examination of ultrasonographic-guided needle-core biopsy specimens (97 dogs), surgical biopsy specimens (85 dogs, which consisted of wedge, needle-core, punch, or excisional biopsy specimens), and 65 specimens obtained during necropsy. On the basis of the histologic reports, there were 85 dogs with malignant disease and 162 dogs with benign disease. Of the malignant diseases, there were 45 primary hepatic neoplasms, 10 round-cell neoplasms, and 30 metastatic neoplasms. The specific neoplastic types and numbers were summarized (Table 1). Benign diseases were not categorized; however, nodular regeneration was detected in 58 dogs and was present in a subset of 12 dogs with malignant disease.
Number and type of malignant tumors determined by histologic examination in 85 dogs with FLLs attributable to malignant diseases.
| Presumptive classification | Tumor type | No. |
|---|---|---|
| Primary hepatic origin | Hepatocellular carcinoma | 35 |
| Cholangiocellular carcinoma | 3 | |
| Hemangiosarcoma | 2 | |
| Sarcoma (specific type not recorded) | 2 | |
| Biliary adenocarcinoma | 1 | |
| Cystadenocarcinoma | 1 | |
| Leiomyosarcoma | 1 | |
| Subtotal | 45 | |
| Metastatic | Hemangiosarcoma | 16 |
| Neuroendocrine (2 classified as pancreatic origin) | 4 | |
| Adenocarcinoma (type not recorded) | 2 | |
| Carcinoma (type not recorded) | 2 | |
| Pancreatic exocrine carcinoma | 2 | |
| Gastrointestinal stromal tumor | 1 | |
| Pheochromocytoma | 1 | |
| Sarcoma (type not given) | 1 | |
| Splenic leiomyosarcoma | 1 | |
| Subtotal | 30 | |
| Round cell | Lymphosarcoma | 5 |
| Histiocytic neoplasm | 5 | |
| Subtotal | 10 |
The distribution of missing test groups by development versus validation data sets and by malignant versus benign histologic results were obtained (Table 2). There was not a significant effect for missing values between the development and validation data sets, except for those of thoracic radiography, which were strongly associated with a malignant histologic outcome. The distribution of candidate predictors in the development and validation data sets was obtained (Table 3). Missing values were most frequent in the test categories of thoracic radiography (up to 55% missing values for the variables abnormal liver and decreased peritoneal detail in the validation data set) and laboratory tests (up to 35% missing values for hypoalbuminemia in the development data set). Duration of clinical signs was the only missing variable in the historical data. There were no missing values for signalment or ultrasonographic data.
Distribution of missing test results by development versus validation data set and by malignant versus benign histologic outcome as determined by evaluation of records and archived diagnostic images for 247 dogs examined at the University of Minnesota Veterinary Medical Center from 2005 to 2008 that underwent abdominal ultrasonography and liver histologic examination.
| Test | Development data set (n = 174) | Validation data set (n = 73) | Malignant (n = 85) | Benign (n = 162) | P value |
|---|---|---|---|---|---|
| Biochemical analysis | 58 (0.33) | 19 (0.26) | 25 (0.29) | 52 (0.32) | 0.77 |
| CBC | 66 (0.38) | 24 (0.33) | 27 (0.32) | 63 (0.39) | 0.33 |
| Coagulation panel | 56 (0.32) | 20 (0.27) | 26 (0.31) | 59 (0.36) | 0.57 |
| Thoracic radiography | 77 (0.44) | 37 (0.51) | 25 (0.29) | 89 (0.55) | < 0.001 |
Data are reported as No. of dogs (proportion with missing data).
Data were analyzed via multivariable logistic regression models. Association between counts of missing test results and histologic outcome was assessed by use of a Fisher exact test; values of P < 0.05 were considered significant.
Distribution of candidate predictors for FLLs with a histologic outcome of malignant versus benign liver disease.
| Candidate predictor | Development (n = 174) | Validation (n = 73) | ||
|---|---|---|---|---|
| No. of dogs (proportion) | Proportion with missing data | No. of dogs (proportion) | Proportion with missing data | |
| Age (y)* | 11.5 (7.9–12.8) | 0 | 11.0 (7.0–14.0) | 0 |
| Sex | ||||
| Neutered male | 75 (0.43) | 0 | 32 (0.44) | 0 |
| Neutered female | 86 (0.49) | 0 | 38 (0.52) | 0 |
| Sexually intact male | 10 (0.06) | 0 | 2 (0.03) | 0 |
| Sexually intact female | 3 (0.02) | 0 | 1 (0.01) | 0 |
| Duration of clinical signs (d)* | 14 (1–365) | 0.09 | 21 (3–240) | 0.03 |
| Vomiting within 10 d prior to abdominal ultrasonography | 50 (0.29) | 0 | 25 (0.34) | 0 |
| Diarrhea within 10 d prior to abdominal ultrasonography | 24 (0.14) | 0 | 15 (0.21) | 0 |
| Appetite decrease within 10 d prior to abdominal ultrasonography | 76 (0.44) | 0 | 34 (0.47) | 0 |
| Medications affecting liver | 19 (0.11) | 0 | 10 (0.14) | 0 |
| Peripheral lymphadenopathy | 11 (0.06) | 0 | 6 (0.08) | 0 |
| Palpable abdominal organomegaly | 23 (0.13) | 0 | 13 (0.18) | 0 |
| Ascites | 2 (0.01) | 0 | 0 (0.00) | 0 |
| Jaundice | 5 (0.03) | 0 | 4 (0.05) | 0 |
| Fever (rectal temperature < 38.9°C) | 46 (0.26) | 0 | 24 (0.33) | 0 |
| Pulmonary nodules | 3 (0.02) | 0.44 | 2 (0.03) | 0.51 |
| Abnormal liver | 43 (0.25) | 0.49† | 18 (0.25) | 0.55† |
| Decreased peritoneal detail | 23 (0.13) | 0.46† | 4 (0.05) | 0.55† |
| Hypoalbuminemia | 53 (0.30) | 0.35 | 23 (0.32) | 0.26 |
| Hyperbilirubinemia | 51 (0.29) | 0.34 | 27 (0.37) | 0.26 |
| High ALP activity | 91 (0.52) | 0.33 | 41 (0.56) | 0.26 |
| High ALT activity | 72 (0.41) | 0.34 | 37 (0.51) | 0.26 |
| Prolonged coagulation time | 38 (0.22) | 0.32 | 24 (0.33) | 0.27 |
| FLL size (cm)* | 2.1 (1–7.8) | 0 | 2.2 (0.8–9.9) | 0 |
| No. of FLLs | ||||
| 1 | 36 (0.21) | 0 | 15 (0.21) | 0 |
| 2 to 5 | 33 (0.19) | 0 | 13 (0.18) | 0 |
| > 5 but discrete | 101 (0.58) | 0 | 42 (0.58) | 0 |
| Complete effacement | 4 (0.02) | 0 | 3 (0.04) | 0 |
| Size variation of FLLs | ||||
| Solitary or ≤ 50%‡ | 60 (0.34) | 0 | 25 (0.34) | 0 |
| > 50%‡ | 114 (0.66) | 0 | 48 (0.66) | 0 |
| FLL echogenicity | ||||
| Isoechoic | 2 (0.01) | 0 | 0 (0) | 0 |
| Hyperechoic | 24 (0.14) | 0 | 8 (0.11) | 0 |
| Hypoechoic | 76 (0.44) | 0 | 33 (0.45) | 0 |
| Heteroechoic | 69 (0.40) | 0 | 31 (0.42) | 0 |
| Hypoechoic with distal acoustic enhancement | 3 (0.02) | 0 | 1 (0.01) | 0 |
| Heterogeneous liver parenchyma | 135 (0.78) | 0 | 58 (0.79) | 0 |
| Hyperechoic liver parenchyma | 11 (0.06) | 0 | 5 (0.07) | 0 |
| Adrenomegaly | 47 (0.27) | 0 | 17 (0.23) | 0 |
| Abdominal lymphadenopathy | 25 (0.14) | 0 | 17 (0.23) | 0 |
| Abdominal mass | 88 (0.51) | 0 | 24 (0.33) | 0 |
Continuous candidate predictor; results for this variable are reported as median (10th to 90th percentiles).
Missing values may reflect dogs for which a single predictor variable within a test was missing or the entire group of tests was not performed.
Represents the percentage difference between the diameters of the smallest and largest FLLs.
Regression coefficients and 95% CIs were obtained for the full, ultrasonographic, and clinical models following backward stepwise selection (Table 4). Decreased appetite, palpable abdominal mass, presence of pulmonary nodules, and elevated ALP or ALT activity were selected via a majority method for the clinical model. None of these variables was associated with a significant regression coefficient, with 95% CIs including the value 0. For the ultrasonographic model, size of the FLL, presence of peritoneal fluid, and a hyperechoic liver were selected. Increased size of FLLs and the presence of peritoneal fluid were associated with malignancy, with positive regression coefficients with 95% CIs excluding the value 0. For the full model, decreased appetite, presence of pulmonary nodules, low albumin concentration, elevated ALP or ALT activity, thrombocytopenia, size of the FLL, presence of peritoneal fluid, and a hyperechoic liver were selected. Increased size of FLLs and presence of peritoneal fluid were associated with malignancy. The c-index results indicated the highest performance for the full model, followed by the ultrasonographic and clinical models. The R2 values indicated that the full model had the most improvement in prediction over the null model (model without any predictor variables), followed by the ultrasonographic and clinical models.
Regression coefficients and 95% CIs obtained for 3 models that incorporated selected candidate predictor variables from all patient information (signalment data, owner-reported medical history, physical examination findings, laboratory test results, thoracic radiographic findings, and abdominal ultrasonographic findings; full model), only the abdominal ultrasonographic findings (ultrasonographic model), or all categories of patient information except for abdominal ultrasonographic findings (clinical model) for predictors selected by use of a majority method and backward stepwise selection on multiple-imputed development data sets (n = 174 dogs in each data set).
| Variable | Full | Ultrasonographic | Clinical |
|---|---|---|---|
| Candidate predictor variable | |||
| Decreased appetite | 0.70 (−0.16 to 1.55) | NA | 0.51 (−0.21 to 1.24) |
| Palpable abdominal mass | NA | NA | 0.89 (−0.13 to 1.91) |
| Pulmonary nodule | −0.76 (−2.06 to 0.55) | NA | −0.94 (−2.14 to 0.26) |
| Low albumin concentration | −0.71 (−1.92 to 0.49) | NA | NA |
| High ALP activity | −1.143 (−2.60 to 0.31) | NA | −1.44 (−2.75 to −0.14) |
| High ALT activity | 0.74 (−0.59 to 2.07) | NA | 0.97 (−0.17 to 2.11) |
| Thrombocytopenia | 0.63 (−0.84 to 2.10) | NA | NA |
| FLL size | 0.26 (0.12 to 0.41) | 0.28 (0.15 to 0.42) | NA |
| Peritoneal fluid | 1.00 (0.16 to 1.83) | 0.96 (0.23 to 1.69) | NA |
| Hyperechoic liver | −1.15 (−2.97 to 0.66) | −1.33 (−3.04 to 0.39) | NA |
| Constant | −1.63 (−2.84 to −0.42) | −1.96 (−2.59 to −1.32) | −0.37 (−1.33 to 0.59) |
| Performance index | |||
| c-index | 0.820 | 0.786 | 0.710 |
| R2 | 0.368 | 0.261 | 0.176 |
Values reported for candidate predictor variables are β (95% CI); β is the regression coefficient for a given candidate predictor variable. Performance index values are reported for each of the 3 models.
NA = Not assessed for the specific multivariable logistic regression model.
Results for model performance against the validation data set were obtained (Table 5). The calibration slope and intercept for the clinical model (0.399 and −0.307, respectively) were farthest from the ideal values (a slope of 1 and intercept of 0 for perfect calibration35). The low value for the calibration slope suggested that the estimated regression coefficients may have been too extreme. Calibration slopes and intercepts for the full and ultrasonographic models also diverged from the values for perfect calibration, but there was less divergence than for the clinical model. All 3 models (full, ultrasonographic, and clinical) also had slopes < 1, which indicated that the estimated regression coefficients may have been too extreme. The c-index values for the full and ultrasonographic models were similar, whereas the c-index value for the clinical model was lower. These results indicated that predictive performance for the full and ultrasonographic models was similar, whereas the clinical model had less predictive performance against the validation data set. Analysis of the R2 values also indicated that incorporation of the selected predictor variables in the full and ultrasonographic models (0.276 and 0.237, respectively) led to an improvement in model prediction against the null model, whereas the R2 value of the predictor variables for the clinical model (0.044) was less helpful in prediction.
Predictive performance for the full, ultrasonographic, and clinical models as determined by use of multiple-imputed validation data sets (n = 73 dogs in each data set).
| Variable | Full | Ultrasonographic | Clinical |
|---|---|---|---|
| Calibration slope | 0.876 | 0.894 | 0.399 |
| Calibration intercept | −0.079 | −0.102 | −0.307 |
| c-index | 0.768 | 0.766 | 0.612 |
| R2 | 0.276 | 0.237 | 0.044 |
Discussion
In the study reported here, we attempted to determine the categories of noninvasive patient data that were most relevant to the diagnosis of benign or malignant disease by comparing the predictive abilities of multivariable logistic regression models fitted to differing categories of patient data. Models incorporating ultrasonographic predictor variables (full and ultrasonographic models) had higher predictive ability for use in predicting benign versus malignant disease than did the model limited to signalment, owner-reported medical history, physical examination findings, radiographic findings, and laboratory test results (clinical model), which suggested that the most relevant information for making the diagnosis was the ultrasonographic features of the FLLs. Indeed, the similarity of the model predictive performance between the full and ultrasonographic models against the validation data set suggested that only the ultrasonographic variables were relevant for predictions against this data set. Only the size of the FLLs and the presence of peritoneal fluid were significantly associated with the histologic outcome.
The single ultrasonographic variable that was most likely to have the greatest effect on model prediction was size of the FLL. In the full model, the estimated logarithmic odds of a malignant disease via histologic examination was 0.28/cm, which corresponded to a 1.3-to-1 odds of malignancy for each 1-cm increase in size of an FLL. The ultrasonographically detected presence of peritoneal fluid was also an influential variable, with an increase in logarithmic odds for malignancy of 1.0 with the detection of fluid, which corresponded to 2.7-to-1 odds of malignancy. In the entire data set, 34 dogs had an HCC; 33 of the HCCs were > 3 cm in diameter, which meant that the effect of size on model prediction approached or exceeded that of peritoneal fluid (logarithmic odds > 0.84 vs 1.0 for the presence of peritoneal fluid). The presence of large HCCs undoubtedly contributed to the prediction of malignancy on the basis of size. Of the remaining neoplasms, 22 were > 3 cm in diameter, and 29 were < 3 cm in diameter. In contrast, there were 40 benign lesions > 3 cm in diameter and 123 benign lesions < 3 cm in diameter. Therefore, the improved predictive ability of models that incorporated ultrasonographic variables, although exceeding that of the clinical model, was partially based on the prediction of malignancy in dogs with large HCCs. Another study26 conducted to investigate the association of ultrasonographic features with liver aspirates found associations between the presence of a large hepatic mass, ascites, abnormal hepatic nodes, and splenic lesions with malignant results on cytologic examination. In the present study, FLL size and peritoneal fluid were predictor variables that had significant associations with malignancy. An association between malignancy and abnormal hepatic nodes was not found. Splenic lesions were not assessed separately from other abdominal masses; however, the presence of an abdominal mass was not associated with malignant results on histologic examination. In a more recent study,d investigators reported marked variability in the ultrasonographic appearance of histologically confirmed liver lesions (both focal and diffuse). Therefore, the lack of association of the echogenicity of FLLs in the present study agrees with findings in other reports. Analysis of our results suggested that the same ultrasonographic variables provided the most information regarding the histologic diagnosis, but this may have been attributable in part to the number of HCCs in the data set. The predictor variable hyperechoic liver (liver echogenicity greater than echogenicity of the spleen) was selected by use of a majority method but was not significant.
Model prediction on the basis of the nonultrasonographic predictor variables was poor, as indicated by the poor performance of the clinical model against the validation data set. None of the predictor variables selected from the categories of signalment, owner-reported medical history, physical examination findings, laboratory test results, or thoracic radiographic findings were significantly associated with outcome. Owner-reported decrease in appetite had a positive nonsignificant (P = 0.11) association with malignancy (95% CI for R2, −0.16 to 1.55). Increased ALP activity had a negative nonsignificant (P = 0.12) association with malignancy (95% CI for R2, −2.60 to 0.31). The negative association for increased ALP activity was likely attributable to the high sensitivity of ALP activity as a predictor for hepatobiliary disease in general, combined with the fact that nonmalignant diseases, such as cholestasis, corticosteroid hepatopathy, chronic hepatitis, and hepatic necrosis, can lead to higher ALP activity than will most hepatic neoplasms.38 These diseases were included in the study data set, which probably led to the higher proportion of dogs with benign diseases with increased ALP activity, compared with the proportion of dogs with malignant diseases with increased ALP activity. The lack of association between pulmonary nodules, which are highly suggestive of metastatic neoplasia, and malignant hepatic disease via histologic examination may have reflected the low number of pulmonary nodules in the data set (3 dogs in the malignant group and 2 dogs in the benign group). It is also possible that pulmonary metastases were not associated with liver neoplasia.
The present study had limitations that need to be considered. The amount of missing data was a severe problem. For owner-reported medical history and physical examination findings, the absence of an entry was equated with the absence of that finding. We believe this approach was justified in that patient medical records often include only abnormal findings. However, abnormal findings for the medical history or physical examination may not have been recorded, which is a limitation of retrospective data. Laboratory tests and thoracic radiographs were not required for inclusion, which also led to a large number of records with missing data. Because we could not assume that missing results were within the reference range or were not abnormal findings, we used multiple imputation techniques to limit the effect of missing data on the analysis. Multiple imputation techniques preserve the relationship of the predictor variables with the outcome variable by filling in missing data by sampling from distributions on the basis of the available data.34 This technique preserves the relationship of the laboratory and radiographic data with the outcome. Although we cannot prove that there was a random pattern for the missing data, there was no statistical association between missing values for laboratory data and outcome. However, a significant (P < 0.001) association was found between missing values for thoracic radiography and histologic outcome, with missing thoracic radiographs associated with a benign histologic outcome. Timing of the thoracic radiography with respect to the abdominal ultrasonographic examination was not established; therefore, it is possible that the abdominal ultrasonographic findings influenced the decision to obtain thoracic radiographs for evaluation of metastasis. For instance, dogs with a single large FLL may have been considered better surgical candidates and underwent thoracic radiography. The association of missing thoracic radiographs with benign histologic outcome had the potential to bias the association of thoracic radiographic variables toward an association with malignant disease; however, we did not find a significant association. The fact that the ultrasonographic data were obtained from interpretation of archived still images may have limited the radiologists' ability to detect subtle lesions that may have been more apparent in real time.
Finally, the use of histologic results as the definitive diagnosis and as an inclusion criterion led to a number of biases. First, the retrospective nature of the study prevented confirmation that a lesion detected ultrasonographically was the same lesion from which a sample was collected for histologic examination; therefore, the histologic results were regarded as representative for the FLLs. Additionally, accuracy of the diagnosis depends on the skills of the people who obtained the tissue samples, whether via ultrasonographic-guided biopsy or surgical biopsy or during necropsy. We believe this study reflects the limitations of clinical practice, whereby the accuracy of lesion sampling cannot be confirmed in most cases and the histologic results are considered representative of the entire organ. However, we acknowledge the fact that a particular patient may have had multiple disease processes and that the histologic results may not have reflected the disease process that caused an FLL. Second, the requirement for tissue samples for histologic examination probably led to bias in diagnostic testing because samples were more likely to be obtained from certain lesions by use of needle-core biopsy or surgical wedge biopsy. These included large single lesions that were potentially amenable to surgery or lesions in which benign, medically manageable disease was suspected. This bias was likely partially offset by the inclusion of specimens obtained during necropsy, which likely consisted of more dogs with more severe disease with a poorer prognosis.
For the study reported here, we concluded that the ultrasonographic data provided the most information for predicting benign versus malignant causes of FLLs as determined on the basis of comparisons between the predictive ability of models that incorporated ultrasonographic variables with the predictive ability of a model that incorporated nonultrasonographic predictor variables. Similar to results in other reports, ultrasonographically measured lesion size and ultrasonographically detected peritoneal fluid were positively associated with malignancy; however, we were unable to detect other useful features for the prediction of histologically benign versus malignant disease associated with FLLs.
ABBREVIATIONS
| ALP | Alkaline phosphatase |
| ALT | Alanine transferase |
| CI | Confidence interval |
| FLL | Focal liver lesion |
| HCC | Hepatocellular carcinoma |
Veterinary Diagnostic Services, Marshfield Clinical Laboratories, Marshfield, Wis.
R, version 2.12.2, R Foundation for Statistical Computing, Vienna, Austria. Available at: www.r-project.org. Accessed Sep 30, 2010.
Microsoft Office Excel, Microsoft Corp, Redmond, Wash.
Warren-Smith CMR, Andrew SS, Lamb CR. Lack of associations between ultrasonographic appearance of the liver and histologic diagnosis (abstr). Vet Radiol Ultrasound 2010;51:583.
References
1 Saunders HM. Ultrasonography of abdominal cavitary parenchymal lesions. Vet Clin North Am Small Anim Pract 1998; 28:755–775.
2 Nyland TG, Mattoon JS, Herrgesell EJ, et al. Liver. In: Nyland TG, Mattoon JS, eds. Small animal diagnostic ultrasound. Philadelphia: Saunders, 2002;93–127.
3 Selcer BA. The liver and gallbladder. In: Mahaffey MB, Hudson JA, Selcer BA, et al, eds. Practical veterinary ultrasound. Philadelphia: Williams & Wilkins, 1995;88–106.
4 d'Anjou M. Liver. In: Penninck DG, d'Anjou M, eds. Atlas of small animal ultrasonography. Ames, Iowa: Blackwell Publishing, 2008;217–261.
5 Rothuizen J, Twedt DC. Liver biopsy techniques. Vet Clin North Am Small Anim Pract 2009; 39:469–480.
6 Faverzani S, Chinosi S, Valenti P, et al. Comparison between ultrasonography and cytology of liver focal lesions and parenchyma in the dog and cat. Vet Res Commun 2006; 30(suppl 1): 293–296.
7 Patnaik AK, Hurvitz AI, Liebermann PH. Canine hepatocellular carcinoma. Vet Pathol 1981; 18:427–438.
8 Patnaik AK, Hurvitz AI, Liebermann PH. Canine bile duct carcinoma. Vet Pathol 1981; 18:439–444.
9 Nyland TG. Ultrasonic patterns of canine hepatic lymphosarcoma. Vet Radiol 1984; 25:167–172.
10 Stowater JL, Lamb CR, Schelling SH. Ultrasonographic features of canine hepatic nodular hyperplasia. Vet Radiol 1990; 31:268–272.
11 Cuccovillo A, Lamb CR. Cellular features of sonographic target lesions of the liver and spleen in 21 dogs and a cat. Vet Radiol Ultrasound 2002; 43:275–278.
12 Cartee RE. Diagnostic real time ultrasonography of the liver of the dog and cat. J Am Anim Hosp Assoc 1981; 17:731–737.
13 Nyland TG, Park RD, Lattimer JC, et al. Gray-scale ultrasonography of the canine abdomen. Vet Radiol 1981; 22:220–227.
14 Nyland TG, Park RD. Hepatic ultrasonography in the dog. Vet Radiol 1983; 24:74–83.
15 Feeney DA, Johnston GR, Hardy RM. Two-dimensional, grayscale abdominal ultrasonography for assessment of hepatic and splenic neoplasia in the dog and cat. J Am Vet Med Assoc 1984; 184:68–81.
16 Whiteley MB, Feeney DA, Whiteley LO, et al. Ultrasonographic appearance of primary and metastatic canine hepatic tumors: a review of 48 cases. J Ultrasound Med 1989; 8:621–630.
17 Majewski A, Gratz KF, Brolsche C, et al. Sonographic pattern of focal nodular hyperplasia of the liver. Eur J Radiol 1984; 4:52–57.
18 Rogers JV, Mack LA, Freeny PC, et al. Hepatic focal nodular hyperplasia: angiography, CT, sonography, and scintigraphy. AJR Am J Roentgenol 1981; 137:983–990.
19 Lamb CR, Hartzband LE, Tidwell AS. Ultrasonographic findings in hepatic and splenic lymphosarcoma in dogs and cats. Vet Radiol 1991; 32:117–120.
20 Jacobson LS, Kirberger RM, Nesbit JW. Hepatic ultrasonography and pathological findings in dogs with hepatocutaneous syndrome: new concepts. J Vet Intern Med 1995; 9:399–404.
21 Farrar ET, Washabau RJ, Saunders HM. Hepatic abscesses in dogs: 14 cases (1982–1994). J Am Vet Med Assoc 1996; 208:243–247.
22 Schwarz LA, Penninck DG, Leveille-Webster C. Hepatic abscesses in 13 dogs: a review of the ultrasonographic findings, clinical data and therapeutic options. Vet Radiol Ultrasound 1998; 39:357–365.
23 Cruz-Arambulo R, Wrigley R, Powers B. Sonographic features of histiocytic neoplasms in the canine abdomen. Vet Radiol Ultrasound 2004; 45:554–558.
24 O'Brien RT, Iani M, Matheson J, et al. Contrast harmonic ultrasound of spontaneous liver nodules in 32 dogs. Vet Radiol Ultrasound 2004; 45:547–553.
25 Scharf G, Deplazes P, Kaser-Hotz B, et al. Radiographic, ultrasonographic, and computed tomographic appearance of alveolar echinococcus in dogs. Vet Radiol Ultrasound 2004; 45:411–418.
26 Guillot M, d'Anjou M, Alexander K, et al. Can sonographic findings predict the results of liver aspirates in dogs with suspected liver disease? Vet Radiol Ultrasound 2009; 50:513–518.
27 Bree RL, Schwab RE, Neiman HL. Solitary echogenic spot in the liver: is it diagnostic of a hemangioma? AJR Am J Roentgenol 1983; 140:41–45.
28 Eberhardt SC, Choi HP, Bach AM, et al. Utility of sonography for small hepatic lesions found on computed tomography in patients with cancer. J Ultrasound Med 2003; 22:335–343.
29 Magne ML, Withrow SJ. Hepatic neoplasia. Vet Clin North Am Small Anim Pract 1985; 15:243–256.
30 Kruth SA, Carter RF. Laboratory abnormalities in patients with cancer. Vet Clin North Am Small Anim Pract 1990; 20:897–917.
31 Strombeck DR, Guilford WG. Hepatic neoplasms. In: Guilford WG, Center SA, Strombeck DR, et al, eds. Strombeck's small animal gastroenterology. 3rd ed. Philadelphia: WB Saunders Co, 1996;847–859.
32 Center SA. Interpretation of liver enzymes. Vet Clin North Am Small Anim Pract 2007; 37:297–333.
33 Cullen JM. Summary of the World Small Animal Veterinary Association Standardization Committee guide to classification of liver disease in dogs and cats. Vet Clin North Am Small Anim Pract 2009; 39:395–418.
34 van Buuren S, Groothuis-Oudshoorn K. MICE: multivariate imputation by chained equations in R. J Stat Softw 2011; 45(3): 1–67.
35 Vergouwe Y, Royston P, Moons KGM, et al. Development and validation of a prediction model with missing predictor data: a practical approach. J Clin Epidemiol 2010; 63:205–214.
36 Rubin DB. Multiple imputation for nonresponse in surveys. New York: Wiley, 1987.
37 Burd RS, Jang TS, Nair SS. Predicting hospital mortality among injured children using a national trauma database. J Trauma 2006; 60:792–801.
38 Center SA, Slater MR, Manwarren T, et al. Diagnostic efficacy of serum alkaline phosphatase and γ-glutamyltransferase in dogs with histologically confirmed hepatobiliary disease: 270 cases (1980–1990). J Am Vet Med Assoc 1992; 201:1258–1264.
