Register for free to
view the full innovation profile.
The innovation removes the artifacts and reflections from the acquired skin images, enhances the image and automatically adjust contrast and brightness of the image.
About
In recent years, there has been a noticeable increase in skin cancer cases, with projections indicating an exponential growth. This innovation introduces a computer-aided diagnosis system designed for classifying malignant lesions. The acquired images undergo primary preprocessing using innovative techniques, eliminating digital artifacts such as hair follicles and blood vessels. Subsequently, an enhancement process employs novel histogram equalization methods to improve image quality. Following this, a segmentation phase utilizes the Neutrosophic technique, employing a threshold-based approach alongside a pentagonal neutrosophic structure to form a segmentation mask for suspected skin lesions.The innovation proposes a deep neural network based on Inception and residual blocks, integrating a softmax block after each residual block to widen the layer and expedite the learning of key features. The proposed classifier undergoes training, testing, and validation across PH2, ISIC 2017, ISIC 2018, and ISIC 2019 datasets. The segmentation model achieves accuracy scores of 99.50%, 99.33%, 98.56%, and 98.04% across these respective datasets. Augmentation of these datasets results in a total of 103,554 images for training, enhancing the classifier's classification accuracy. Our experimental results affirm the superiority of the proposed classifier, outperforming most existing classifiers with accuracy scores of 99.50%, 99.33%, 98.56%, and 98.04% for PH2, ISIC 2017, 2018, and 2019 datasets.
Key Benefits
This innovation presents an effective pre-processing model designed to remove image artifacts and enhance the image through histogram equalization. Following refinement, the image undergoes segmentation using a mathematical-based algorithm involving thresholding and pentagonal neutrosophy, achieving superior segmentation results. The segmented image is then classified using our proposed classifier trained with a large augmented and balanced dataset.Evaluation against publicly accessible datasets—PH2, ISIC 2017, 2018, and 2019—demonstrated the superior performance of our proposed methods in various stages of computer-aided diagnosis (CAD) over state-of-the-art methods. Notably, our method achieved significantly higher sensitivity and specificity in diagnosing melanoma lesions, surpassing other classifiers like YOLO, KNN, and Bayesian networks.The importance of our proposed techniques in data augmentation, pre-processing, and segmentation is evident from the comparison. Our deep neural network outperformed models employing only standard augmentation methods. Specifically, employing our proposed techniques increased sensitivity and specificity scores by 8–10%.Future research aims to explore diverse datasets encompassing multiple skin diseases and various skin cancers for more comprehensive classifications. We plan to deploy these algorithms into smartphone applications, offering skin conditions monitoring and non-invasive skin cancer diagnosis. Moreover, we aim to introduce a smartphone-based dermoscopic tool to enhance diagnosis accuracy. The digital diagnosis of skin lesions holds vast potential for future advancements
Applications
The innovation could be used for monitoring skin conditions by health sectors and beauty companies.