AI calls the bluff: differentiating benign lesions from triple-negative breast cancer cases.
Authors
Affiliations (5)
Affiliations (5)
- Instituto de Biofísica e Engenharia Biomédica, Faculdade de Ciências da Universidade de Lisboa, 1749-016, Lisbon, Portugal. [email protected].
- LASIGE, Faculdade de Ciências da Universidade de Lisboa, 1749-016, Lisbon, Portugal. [email protected].
- Instituto de Biofísica e Engenharia Biomédica, Faculdade de Ciências da Universidade de Lisboa, 1749-016, Lisbon, Portugal.
- Hospital de Vila Franca de Xira, Estrada Carlos Lima Costa Nº2, 2600-009, Vila Franca de Xira, Lisbon, Portugal.
- LASIGE, Faculdade de Ciências da Universidade de Lisboa, 1749-016, Lisbon, Portugal.
Abstract
Triple-negative breast cancer (TNBC) is the most aggressive molecular subtype of breast cancer (BC). TNBC lacks targeted treatment options, which results in poor clinical outcomes. TNBC lesions usually present benign characteristics on mammograms, complicating their early diagnosis. This retrospective multicenter study presents a convolutional neural network (CNN) model to distinguish TNBC from benign lesions on 566 mammograms (277 benign/289 TNBC), acquired at three different institutions across the UK. Each mammogram had its quality enhanced using a combination of total variation minimization filtering and contrast local adaptive histogram equalization (CLAHE). The proposed model achieved a test set AUC of 0.984, with a sensitivity and specificity of 94.2% and 91.9%, respectively. Explainability with GRAD-CAM was applied to the test set, revealing that the model was using not only lesion characteristics but also tumor microenvironment regions to make predictions. The same test set was analyzed by an expert radiologist who achieved a sensitivity of 71% and a specificity of 60%. The comparison of results between the developed model and the expert radiologist highlights the model's performance and underscores its potential as a complementary diagnostic tool. This model might help in the task of TNBC early diagnosis, potentially diminishing the number of false negatives.