A Comprehensive Study of Mammogram Classification Techniques
Parita Oza, Yash Shah, Marsha Vegda
Abstract
Cancer instances have increased in the recent past and are risking many lives. Every kind of cancer is caused due to a malignant (cancerous) tumor which looks pretty similar to a benign (non-cancerous) tumor which makes it difficult to distinguish from one another. Computer-Aided Diagnosis (CAD) is very useful in the early detection of a tumor from its development stage. Many techniques are available in the literature for lesion or mass classification but the challenges faced to train a model are also of great concern. There comes a question about which method to implement for early detection of cancer. In this paper, we have discussed various classification techniques that are categorized based on function, probability, rule and similarity. Analysis of these methods, their drawbacks, and challenges are also discussed. Comparative analysis of existing approaches used to classify mammograms is also presented in the paper. The challenges to train a model also have come from the type of dataset used for classification process. Sometimes the dataset has many anomalies like redundancy, variable size and inconsistency in dimensions that cannot be negotiated. Sometimes the model (architecture) we chose for the training purpose has extensive computation, making it inefficient. These challenges have to be solved during the preprocessing and training phase of a model. This paper mainly focused at various challenges related to mammogram datasets and classification techniques. Methods to handle these challenges are also discussed in the paper.