Multi-label classification analysis with modified C-Tran on SCIN dataset ,

Skin conditions affect millions of people globally, with symptoms appearing in different body areas. Technological advancements have brought diverse data types, including situations where an image depicting a skin condition can be assigned multiple labels. The Classification Transformer (C-Tran) method, which utilizes transfer learning and transformers, was developed for multi-label classification. Recently, Google introduced a new dataset called SCIN (Skin Condition Image Network), which aims to provide diverse data on skin conditions. This research aimed to use the C-Tran method for the multi-label classification of skin conditions with the SCIN dataset while incorporating additional metadata inputs to improve the metric results. The results show that the multi-label classification process using metadata is far superior to the model without metadata. For example, In the mAP metric, models that utilized metadata scored 82.37, whereas models without metadata only scored 47.02. Similarly, models with metadata achieved 70.83% in the accuracy metric, while models without metadata achieved only 34.72%. Out of the 10,379 data points available with metadata in the SCIN dataset, only 718 were actually utilized for the classification task. It is thought that the inaccurate prediction outcomes are due to unreliable data, even with a confidence level of 4. In this analysis, two metadata categories stood out the most in terms of different measurements: the body part and symptoms metadata categories from the SCIN dataset. With just the body part and symptoms metadata groups, the mAP results achieved a 74.23%, accuracy at 63.89%, CF1 at 68.79%, and OF1 at 73.13%.