A FUSED PARALLEL RESNET MODEL FOR ENHANCED SKIN CANCER DETECTION

Abstract

To achieve high survivability quotient, there is the need to have precise diagnostic procedures that will help in early cancerous skin detection as it remains one of the problems of life taken by cancer all over the world. The visual jobs by dermatologists in clinical diagnosis are found to be subjective and time consuming in examinations. This paper propose a deep learning model based on dual-branch ResNet50 and ResNet100 to help handle the problem of poor performance in both feature extraction and classification. In TensorFlow, the ImageDataGenerator component serves the purpose of preparing data, as well as its improvement prior to the distribution of the dataset into training and validation. The feature representations of the two ResNet networks are pretrained through ImageNet prior to concatenation of the vectors so that they can be processed through full-connected layers and other groups such as batch normalization and global average pooling of the information until they are subjected to encoding through softmax multi-class classification. Checkpointing of the system can allow the continuity of training and system resilience. ResNet50-ResNet101 exhibited the high predictive consistency by lowering biases along with the increased robustness to data variations that results in the higher accuracy of detection across multiple types of lesions. The quality of the method is measured in terms of Accuracy, Brier Score and Cross-Validation Score as well as Grad-CAM and LIME (Local Interpretable Model-agnostic Explanations) to make it reachable. The detection system demonstrated a high potential in skin cancer detection due to the accuracy of classification to 95%.