Image Processing-based Resource-Efficient Transfer Learning Approach for Cancer Detection Employing Local Binary Pattern Features.

Cancer is a life-threatening disease and has witnessed a substantial increase during the past few years. With a further expected increase in the future, it poses a huge challenge for the medical industry to design accurate detection approaches. Earlier and accurate detection of these cancers increases the surviving chances of patients. Computer-aided diagnosis (CAD) can provide significant help in assisting medical experts to make timely and accurate diagnoses. Deep learning models show great promise but are limited by the need for larger datasets for training and higher computational complexity. This study proposes a resource-efficient EfficienetNetB4 model that incorporates local binary pattern features to enhance the detection accuracy of cancer. The proposed approach integrates histopathological images and texture-based features with EfficientNetB4. Extensive experiments involving several pre-trained convolutional neural network variants are carried out. Results show the superior results of the proposed EfficientNetB4 model with a 99.8% accuracy. In addition, precision, recall, and F1 scores are 99.9% each thereby indicating the exceptional results for cancer detection exceeding the existing state-of-the-art approaches. Using a 5-fold cross-validation provides an average 99.88% accuracy further validating the performance of EfficientNetB4. Furthermore, the use of Shapley additive explanations helps comprehend the results and increases the transparency of the decision-making process. This study potentially contributes to further research in CAD-based resource-efficient cancer detection. [ABSTRACT FROM AUTHOR]