| Popis: |
Skin diseases[1] consists of a wide range of ailments that affect the skin, including microbial infections, viral, fungal, allergies, epidermis malignancies, and parasitic diseases. In South-Asian countries like India people don’t care much about the skin conditions. In our country, people prefer home remedies to cure skin conditions instead of visiting a dermatologist which can lead to serious skin conditions. Early diagnosis of skin disease is very important as it can reduce the severity of the condition. Melanoma [1] is the deadliest type of skin cancer, and it is the most prominent form of cancer. Melanoma could be diagnosed early, which would reduce overall illness and death. The odds of dying from the ailment is proportional to the extent of the malignancy, which is proportional to the length of time it has been growing. The keys to early detection are patient self-examination of the skin, full-body skin screenings by a dermatologist, and patient engagement. This work aims to categorize skin cancer into two types: malignant and benign. Two different approaches were used. Starting with a simple Convolutional Neural Network [2] and then moving on to transfer learning [1]. Image processing, as the name suggests, involves processing the visual, which might comprise a range of techniques until we accomplish our mission. The outcome could either consist of a visual or a component that belongs to that image. This can be employed in decision-making and subsequent assessment. The Dataset we have used in this work consists of train data with 2637 images and test data with 660 images belonging to 2 classes, malignant and benign. This work aims to categorize skin cancer into two types: malignant and benign. Two different approaches are used. Starting with a simple Convolutional Neural Network and then moving on to transfer learning. The major goal is to evaluate the performance of the two models and then see how transfer learning can improve accuracy, particularly when there is not enough data. An accuracy of 91 percent was achieved using the improver module which was better than our previous research where we achieved an accuracy of 82 percent. |
