Search

A computer-aided diagnostic tool for myopia detection is of paramount importance since the pathology results in irreversible damage to the eye. Unfortunately, designing of such systems is hindered by several issues such as (i) difficulties in segmentation of optic disc (OD) due to smooth variability between the OD and other retinal features, (ii) greater degree of correlation between myopia and non-myopic features, and (iii) lack of feasibility for adopting in a clinical setting. The proposed methodology is designed to address the aforementioned issues. First, blood vessels are detected and excluded. Second, the optic disc region is segmented using deformable models. Further, two classification set-ups are tested to determine the rate of accurate classification to detect myopia. Quantitative analysis is performed on the PALM dataset achieving, sensitivity, specificity, and accuracy of 90.3 %, 100 %, and 95 % respectively.

The rapid rise in skin diseases over the past decade has been a growing concern worldwide. Early detection, correct categorization, and accurate identification can result in the successful treatment of melanoma, thereby decreasing the morbidity and mortality rate. Thus, there is a significant need for a system that is capable of identifying skin diseases and precisely classifying them. The proposed work aims to develop a multi class classification system using transfer learning-based convolutional neural networks (CNN). In particular, the proposed solution classifies the dermoscopic images to 8 different categories namely Melanoma (MEL), Basal Cell Carcinoma (BCC), Actinic Keratosis (AK), Benign Keratosis (BKL), Dermatofibroma (DF), Vascular lesions (VASC) and Squamous Cell Carcinoma (SCC). Four state-of-art pre-trained models are used for this task. A functional model-based network is leveraged to embed these sub-models in a larger multi-headed neural network. This will allow the embedded model to be treated as a single large model. An ensemble approach, termed as blending, is employed to combine the predictions efficiently made by the sub-models. Additionally, a robust cropping strategy is implemented to deal with the uncropped images and their impact on the performance of the classifiers is investigated. The impact of applying blending technique to ensemble the pre-trained CNNs are investigated against the performance of the individual classifier. The proposed work is carried out on International Skin Imaging Collaboration (ISIC) 2019 dataset. In this work, the solution for task 1 of the challenge is presented and we obtained balanced multi-class accuracy of 81.2% on the dataset compiled from the original dataset.