Non-Invasive Cancer Detection for Lumps & Bumps Emerges as a New Category of Care for Practitioners...
Background and Clinical Need
HT Vista is a non-invasive cancer detection tool originally developed as a screening tool to help veterinary teams quickly assess whether a dermal or subcutaneous mass is likely benign or requires further diagnostic testing 1. Using AI-powered heat diffusion analysis, HT Vista detects features of malignancy in thermal patterns in tissue.
Efficiently diagnosing lumps and bumps in dogs is crucial for successful treatment. The standard method of cytology is effective but can be costly, invasive, and inconclusive. As a result, many pet owners choose to "wait and see," leaving masses undiagnosed and dogs to go home with cancer.
From a ‘Rule Out’ Screening Tool To Also a “Rule In” Diagnostic Modality
HT Vet is happy to introduce its updated version 4.0 with major improvements to its cancer detection modality. Advancements in AI and HT Vista’s expanding dataset have paved the way for two exciting new features:
-
Cancer Risk Values and Levels: These offer a visual framework to help veterinary teams assess lumps and bumps on the spot, aiding in decision-making by accurately identifying low-risk cases that can be confidently monitored, while guiding next steps for moderate- and high-risk cases.
-
AI-powered Tumor Subclassifiers: Flag Lipomas and Mast Cell Tumors (MCTs) - in cases where HT Vista detects very low or very high malignant probability - minimizing unnecessary invasive procedures for Lipomas and enabling early detection of MCTs.
This article explains the technical details of the new features and highlights the performance of the Lipoma and MCT subclassifiers.
Review: Training Dataset Composition
The primary classifier was trained on a dataset of 1,163 canine masses, with tumor types distributed as in table 1.
Table 1. Distribution of Dataset Across Tumor Types
Validation Methods and Study Design
After training the classifier, a validation study was performed, utilizing a three-step framework for classifying dermal and subcutaneous masses in dogs:
Step 1: Primary Classification - creating a Model Score
For each validated scan, the model outputs a discrete predictive Model Score ranging from 0 to 1. Image 1 shows a spectrum of Model Score results of all 1,163 scans.
The Diagnostic “Ground Truth” - from cytology and / or histopathology - is then charted for each. Image 1 shows the distribution of results (benign vs. malignant) across Model Scores, from which HT Vista's overall Sensitivity and Specificity is derived.
Image 1. Distribution Spectrum of Model Scores with Corresponding Diagnostic Results (n = 1,163).
Step 2: Converting Model Score to Cancer Risk Value
To ensure clinically relevant results, the Model Scores from the primary classifier are adjusted for prevalence 2. The original dataset had a benign-to-malignant ratio of 70:30 (table 1) whereas real-world veterinary clinical data suggests an 85:15 ratio.
Prevalence Adjustment Process:
• The classifier’s Model Scores were translated into Cancer Risk Values using Bayes’ theorem2 to reflect real-world prevalence.
• Cancer Risk Values are categorized into three Cancer Risk Levels: low, moderate or high Cancer Risk, for better clinical decision-making (Image 2).
Image 2: Tumor Type Distribution by Cancer Risk
Step 3: Subclassification Models for Tumor Detection
Only high-confidence cases from the primary classifier were used to train the subclassification models.
|
The Lipoma subclassifier was trained with very low Model Score (≤0.18) cases. 249 of 453 Lipomas in the dataset met these criteria (table 2). |
The MCT subclassifier was trained with very high Model Score (≥0.89) cases. 111 of 229 MCTs in the dataset met these criteria (table 3). |
Validation of Lipoma and MCT subclassifiers
|
The Lipoma subclassifier showed 90% specificity, with a 99% PPV. The sensitivity was 69%, indicating not every Lipoma is flagged (table 4). This suggests if a lump is flagged as Lipoma, it's highly likely, and while not every Lipoma is flagged, a low Cancer Risk Level ‘Rules Out’ malignancy with high confidence. |
The MCT subclassifier showed 90% specificity, 77% PPV. The sensitivity was 44% indicating some MCTs are not flagged (table 4). This suggests if a lump is flagged as MCT, it’s highly likely, and while not every MCT will be flagged, a high Cancer Risk Level ensures malignancy is detected via further investigation. |
Table 2. Lipoma Subclassifier Confusion Matrix*
Table 3. MCT Subclassifier Confusion Matrix*
How can this help you?
Enhance Veterinary Decision-Making
Cancer Risk Levels enable better case prioritization. Low-risk cases can be ruled out with confidence, while moderate- and high-risk cases are flagged for further investigation. Tumor-specific subclassifier flags provide actionable insights, facilitate more informed discussions with pet owners, ultimately helping veterinary teams adopt proactive diagnostic and treatment strategies.
Reduce Unnecessary Invasive Procedures
Combining a low Cancer Risk Level with a Lipoma subclassifier flag reduces the need for invasive diagnostics. This replaces the often adopted “wait and see” approach with more confident, immediate decision-making.
Promote Early Detection
HT Vista enables a proactive approach that increases the likelihood of detecting malignant tumors earlier, enabling more effective treatment and improving patient outcomes, particularly for MCTs, where early intervention can improve prognosis.
Table 4: Performance Evaluation of Primary Classifier and Subclassifiers
Conclusion
HT Vista’s tumor-specific subclassifiers transform it from a ‘Rule Out’ screening tool to a ‘Rule In’ diagnostic modality. The primary classifier provides high sensitivity and NPV for accurate cancer screening, while the subclassifiers offer high specificity and PPV for accurate tumor identification. Cancer Risk Levels further improve decision-making, reducing unnecessary procedures and promoting early detection and intervention. These findings support the integration of HT Vista into veterinary practices and demonstrate its potential for enhancing diagnostic accuracy and patient outcomes.
References
-
Dank G, Buber T, Rice A, Kraicer N, Hanael E, Shasha T, Aviram G, Yehudayoff A, Kent MS. Training and validation of a novel non-invasive imaging system for ruling out malignancy in canine subcutaneous and cutaneous masses using machine learning in 664 masses. Front Vet Sci. 2023 Sep 29;10:1164438. doi: 10.3389/fvets.2023.1164438. PMID: 37841459; PMCID: PMC10570610.
-
Webb MPK, Sidebotham D. Bayes' formula: a powerful but counterintuitive tool for medical decision-making. BJA Educ. 2020 Jun;20(6):208-213. doi: 10.1016/j.bjae.2020.03.002. Epub 2020 Apr 19. PMID: 33456952; PMCID: PMC7808025.
*In tables 2 and 3, you will find the following initials with the following meanings: (TN - True Negative), (FN - False Negative), (TP - True Positive), (FP - False Positive)
