INTELLIGENT SKIN CANCER DIAGNOSIS USING PARTICLE SWARM OPTIMIZATION (PSO) AND TRANSFER LEARNING

Abstract

It's true that when our skin spends a lot of time in the sun, abnormal cell growth can occur, and that's what we know as skin cancer. Interestingly, this common cancer can also pop up on areas of skin that don't see much sunlight. The main types of skin cancer are melanoma, squamous cell carcinoma, and basal cell carcinoma. While skin cancer can be serious, even fatal, the outcome really depends on a few things: the specific type of skin cancer, the person's overall health, and how early the cancer is found. Melanoma is often called "the most serious" because it has a greater tendency to spread. It can develop within an existing mole or suddenly appear as a dark spot that looks different from the surrounding skin. On the other hand, basal cell and squamous cell carcinomas are less likely to be life-threatening.

The good news is that artificial intelligence is making huge strides in quickly and accurately identifying many diseases, which is helping people get treatment sooner. One AI technique, called convolutional neural networks, is particularly good at looking at images and providing very precise information.

One study explored using an efficient convolutional neural network model to analyze skin cancer images. They then used the model to extract key features from these images. To get the best features, they tried out different combinations and used optimization algorithms called PSO and GA to select them. Finally, they used an SVM approach to classify each set of chosen features and aimed for the best possible results. This method achieved an accuracy of 89.17%, showing that it could be quite effective.

Another study found that a model called U-Net++ with densenet201 as its backbone performed even better across various measures like accuracy, F1-score, AUC, iou, and dice values, achieving scores of 94.16%, 91.39%, 99.3%, 96.8%, 77.19%, and 75.47% respectively.

Sadly, about 3.5 million people in the United States alone are diagnosed with skin cancer each year. The chances of survival drop significantly as skin cancer progresses. However, finding this type of cancer early can be difficult and expensive. To address this, one study used an automated, threshold-based method to detect, categorize, and segment skin cancer cases. They even used a smart optimization tool called spasa to fine-tune the settings of eight well-known CNN models (like VGG16, VGG19, mobilenet, and nasnet) to get the best results.