Real-time microscopy image-based segmentation and classification models for cancer cell detection

Image processing techniques and algorithms are extensively used for biomedical applications. Convolution Neural Network (CNN) is gaining popularity in fields such as the analysis of complex documents and images, which qualifies the approach to be used in biomedical applications. The key drawback of the CNN application is that it can't call the previous layer output following the layer's input. To address this issue, the present research has proposed the novel Modified U-Net architecture with ELU Activation Framework (MU-EAF) to detect and classify cancerous cells in the blood smear images. The system is trained with 880 samples, of which 220 samples were utilized in the validation model, and 31 images were utilized to verify the proposed model. The identified mask output of the segmentation model in the predicted mask fits the classification model to identify the cancer cell occurrence in the collected images. In addition, the segmentation evaluation is done by matching each pixel of the ground truth mask (labels) to the predicted labels from the model. The performance metrics for evaluating the segmentation of images are pixel accuracy, dice coefficient (F1-score), and Jaccard coefficient. Moreover, the model is compared with VGG-16 and simple modified CNN models, which have four blocks, each consisting of a convolutional layer, batch normalization, and activation layer with RELU activation function that are implemented and for assessing the same images used for the proposed model. The proposed model shows higher accuracy in comparison.