Computer-Based Segmentation of Cancerous Tissues in Biomedical Images Using Enhanced Deep Learning Model

Abstract

In this manuscript, we proposed an automatic segmentation method which was developed using the depth-wise separable convolution with bottleneck connections. The data were normalized using group normalization for reducing the computational complexities and clipped RELU was used with ceiling capped at 6. The network was trained on the datasets of brain tumor and skin cancer while it was tested on the same as well as different datasets acquired under different environments. Additionally, for the case of the brain tumor, the network was tested on real-time MRI dataset. The quantitative and qualitative analysis of results inferred the superior performance of the proposed network. The mIoU and BF Score were increased by 3% and 4.5% for brain tumor segmentation when the network was tested on the different dataset without retaining. For skin cancer dataset an increment of 3% and 5% was observed in both the evaluation metrics. The results obtained on real-time MRI data of brain tumor showed the improvement of (4.2 ± 0.024)% and (4.6 ± 0.0286)%, respectively, in mIoU and BF score. The proposed model produced accurate boundary and pixel details for medical diagnostic purposes. Experienced radiologists did external validation of the proposed method by comparing the obtained results with the manually segmented images. This computer-assisted approach can save the time and burden of doctors for the diagnosis of cancer.