Implementation of a Method Using Image Sequentialization, Patch Embedding and ViT Encoder to Detect the Breast Cancer on RGBA Images and Binary Masks

This paper uses an approach where we will be training the unlabelled and labelled datasets into the system and these datasets comprises of: RGBA images and Binary masks. This will be a benefit in order to get a much better result and more accurate also as pre-training method always helps in getting excellent output. We will use a Transformer Model in this paper where we will apply an image sequentialization technique having 51 million parameters which will help us in attaining the smoother images without any noise. At times, this approach could be time consuming due to used datasets and its sizes, however it shows more accurate and efficient results. Further, we will try to figure out the pixel wise label map using patch embedding technique. Once these techniques are applied then CNN-Transformer Hybrid will come into the role which will encode and decode the images to high-level feature extractions and full spatial resolution respectively. This way of doing encoding and decoding is also known as forward pass and back propagation. Also, it will involve the cascaded upsampler where we will try to use self-attention processes into the design of encoder using the transformers. This entire mechanism will involve few of the best evaluation metrics and those are: Pixel accuracy, IoU, Mean-IoU and Recall/Precision/F1 Score giving the effective and best results.