Segmentation of microscopic images of dermatophytes in clinical samples using ResNet and attention-based modifications of U-Net

Direct microscopic images of dermatophytes face the challenge of distracting artifacts such as cell debris from skin preparations, air bubbles, cellulose fibers, and also blurred images arising from out-of-focus areas, all of which make the diagnosis at the clinics cumbersome. Automated detection will be the preferred choice under such situations. Work in the area of semantic segmentation of dermatophytes has not yet been attempted, while a few attempts have been reported on the object detection and classification aspects. Our work focusses on the modification of the popular and efficient U-Net architecture with the introduction of residual, squeeze and excitation (SE), and attention-gating modules. A hybrid of weighted binary cross-entropy and dice loss functions was also used to obviate the imbalanced pixel class distribution. The dataset included images captured with 10 × × and 40 × × objective lenses. Significant increase has been progressively noticed in dice score with an appreciable improvement of about 8% and 13% in 10 × × and 40 × × magnifications, respectively. Qualitative performance evaluation conducted using heatmaps and predicted images permitted visual verification of the progress made. All the metrics evaluated gave satisfactory values which indicated better performance of the proposed model over the compared architectures.