Multi-headed CNN for colon cancer classification using histopathological images with tikhonov-based unsharp masking

Colon cancer (CC) is a leading cause of mortality rate across worldwide. Early diagnosis of colon cancer prolongs human life and is also helpful in preventing the disease. Histopathological inspection is a frequently used technique to detect and diagnose it. Visual inspection of histopathological analysis needs more inspection time and the decision is based on clinicians' subjective perceptions. Typically, machine learning techniques depend on conventional feature extraction which is time-consuming and laborious, and may not be suitable for a large amount of data. This work proposed a lightweight multi-headed convolutional neural network (MHCNN) for the classification of colon tissue using histopathological images. Authors have employed the first-time Tikhonov-based unsharp masking on colon tissue histopathological images. Tikhonov-based unsharp masking is used to highlight the high-frequency details (such as edges, corners, contours, etc.) of histopathological images. Then, the obtained unsharp mask-based histopathological images are given as input to the proposed MHCNN. Further, quantized aware training is applied to the proposed model to reduce the model size for efficient storage and speed up the training and testing time while maintaining high accuracy. The effectiveness of the developed model is compared with the other existing state-of-the-art methods. The proposed MHCNN achieved an average classification accuracy of 96.62%, precision of 97.48%, specificity of 97.46%, f1 score of 0.9664, and area under the curve of 0.9828. The quantized accuracy of 96.10% is achieved by the proposed network. Clinicians may install the developed network to validate the diagnosis in the hospitals.