Breast invasive ductal carcinoma detection with histopathological images using deep learning

Voon, Wingates (2022) Breast invasive ductal carcinoma detection with histopathological images using deep learning. Final Year Project, UTAR.

Preview

PDF Download (7Mb) | Preview

Abstract

The staining of haematoxylin and eosin (H&E) in histopathological samples leads to inconsistent colour and intensity variations among digital datasets, thus hindering the performance of deep learning computer-aided diagnostic (CAD) systems. One proposed technique to battle colour invariance among digitalised histopathological images is stain normalisation (SN), which adjusts the source image colour to match the overall colour distribution of other similar images in a dataset. Some studies claimed that SN techniques improved CNNs' performance in histopathological classification tasks, while several contradicted their claims. Therefore, we attempt to justify the importance of SN, specifically Reinhard and Macenko techniques in the invasive ductal carcinoma (IDC) grading application using seven selected CNN models: EfficientNetB0, EfficientNetV2B0-21k, ResNetV1-50, ResNetV2-50, MobileNetV1, and MobileNetV2. Our findings indicated that CNN models trained in the original (non-normalised) dataset outperformed models trained with SN datasets. Among the two SN techniques, the Reinhard average scores topped the Macenko across all evaluation metrics in cross validation (cv) and test results while being more consistent in performance. Hence, we suggest that SN is considered unnecessary to be included in the CNN pre-processing steps to improve CNN performance if effective CNN architectures are employed.

Actions (login required)