Abstract
Identifying mitosis is crucial for cancer diagnosis, but accurate detection remains difficult because of class imbalance and complex morphological variations in histopathological images. To overcome this challenge, we propose a Customized Deep Learning (CDL) model, which integrates advanced deep-learning techniques for better mitosis detection. The CDL model utilizes transfer learning to counter the effects of class imbalance and speed up convergence, while skip connections are also employed to improve the localization of mitosis. Furthermore, we have established an innovative selection mechanism by the hybrid of Jellyfish Search Optimizer (JSO) and Walrus Optimization Algorithm (WOA) to maximize the momentum of the model. The proposed approach is rigorously evaluated on multiple publicly available mitosis detection datasets, including Mitosis WSI CCMCT Training Set, Mitosis-AIC, Mitosis Detection, and Mitosis and Non-Mitosis datasets. To tackle these issues, we hereby bring forth a specifically tailored Custom Deep Learning model, that assimilates hybrid CNN architecture into transfer learning and feature selection for improved mitotic detection. The CDL model comprises a Transfer Learning-based Mitosis Detection module under which extracted features from pre-trained deep networks are used to bolster feature extraction and alleviate class imbalance through skip connections to better localize mitosis. The robust assessment on a benchmark dataset displays the outstanding efficacy of the CDL model, reaching an excellent F1 score of 0.994 and accuracy of 98.8% thus proving its strength for the detection of mitotic figures. This proposed methodology can greatly empower pathologists for accurate appraisal of cancer diagnosis and prognosis. Future lines of exploration will include fusion methodologies and time efficiency for real-time applications, as well as extending CDL to various histopathological analyses.