Evaluating Interpretability in Deep Learning using Breast Cancer Histopathological Images

• Daniel C. Macedo UFPE
• John W. S. De Lima UFPE
• Vinicius D. Santos UFPE
• Tasso L. O. Moraes UFPE
• Fernando M. P. Neto UFPE
• Nicksson Arrais SiDi
• Tiago Vinuto SiDi
• João Lucena SiDi

Resumo

Breast cancer is the most common cancer type and is the leading cause of death among females worldwide. Despite these negative statistics, early diagnosis gives patients a high probability of survival. In literature, diagnostic techniques based on histopathological images were proposed for early diagnosis. However, they are limited because they depend on the pathologist’s work and experience. In other words, a patient may receive a different diagnosis from different pathologists, or inexperienced pathologists may misdiagnose. In this work, we implement five Deep Neural Networks (DNNs) and evaluate classification accuracy and interpretability from real tumour images. To evaluate our models, we propose a metric to assess the interpretability of Deep Neural Networks (DNNs). The experiments with the BreCaHAD annotated dataset have shown that MobileNetV2 presented a higher accuracy in classifying histopathological images and interpreting their features, leading the way to improve the pathologist’s work

Palavras-chave: Deep learning, Training, Malignant tumors, Neural networks, Computer architecture, Transformers, Breast cancer