HTCH-Loss: A Hybrid Loss Function with Correlative Capsule Scale Attention Network for Lung and Colon Cancer Detection using Histopathological Images

Abstract

Lung and colon cancers are among the most common cancer in all over the world, affecting both males and females. Early detection is crucial to reduce mortality. However, the difficulty of diagnosis often depends on the histopathologist’s experience. Histopathological images offer detailed tissue-level information, but manual analysis is time-consuming and labor-intensive. Variability within classes and inconsistent staining make diagnosis challenging. Recently, Deep Learning (DL) has attained popularity for improving medical image interpretation. Therefore, a new module for Lung and Colon Cancer detection from histopathological images named Harmonic Tanimoto constrained Huber loss (HTCH-Loss) with Correlative Capsule Scale Attention Network (CCSAN) framework is designed to tackle these problems. The histopathological images based on minority and majority classes are initially fed to the Dual-Generative Adversarial Network (GAN) mechanism for realistic synthetic image generation. Here, both real and synthetic images are obtained. After that, the histopathological image is integrated to form a new dataset. Then, blood cell segmentation is done by Lightweight Feature Attention Network (LFANet). Thereafter, Deep Ensemble Feature Extraction (Deep EFE) is performed. Afterwards, lung cancer detection is done using CCSAN with HTCH-Loss. Moreover, CCSAN is modified by using HTCH-Loss, which is the integration of Tanimoto loss, constrained focal loss, and Huber loss, with the harmonic concept. Simultaneously, a similar process is followed for colon cancer detection. Experimental results demonstrate that the proposed CCSAN with HTCH-Loss attains superior accuracy values of 95.237% for lung cancer detection and 96.669% for colon cancer detection.