Publication

Abstract : In a number of tasks for assessing natural and medical prints, deep neural networks have bettered humans. These achievements, however, are solely reliant on properly labelled training data. When presented with a few samples of noisy-labelled images, the network training strategy might encounter difficulties, resulting in a suboptimal classification models. The quality of medical picture annotations strongly depends on the knowledge and experience of the annotators, which makes this difficulty more acute in the context of medical image analysis. Such problem arises due to a number of reasons ranging from human error, inexperience or even misreading of the images. But at the same time, proper labels are exceptionally useful in training models, while improperly labelled data actually hampers the efficiency of the model. Label for a lot of the image dataset out there is actually already generated, but when we think about the noisy labelled data and go for an unsupervised learning model instead of a supervised model the labels that are correct are also going to waste. To solve this problem without wasting labels, a sparsely supervised learning strategy based on transfer learning is constructed with the aid of the keras Xception model. This paper compares the efficiency of a sparsely supervised learning model employing transfer learning to that of other traditional CNN models based on their performance metrics. © 2023 IEEE.