MEDICAL IMAGE ANALYSIS FOR DIAGNOSING AND CLASSIFYING BREAST CANCER USING DEEP LEARNING

Abstract

Breast carcinoma is the premier category of deadliest cancer reported in females. Breast cancer cases are rising dramatically both in India and around the world, especially in individuals between the ages of 30 and 40. Automatic diagnosis of breast cancer is necessary because manual diagnosis is laborious and time-consuming. Our current healthcare systems are prone to failure. Late detection is the main cause of low breast carcinoma survival rates in the country. Consequently, computer-aided diagnosis (CAD) for medical imaging has become a useful gadget for physicians to categorize clinical images into several groups, facilitating early diagnosis and treatment. Machine Learning (ML) and Deep Learning (DL) have developed various techniques/algorithms for diagnosing and classifying breast cancer early. Multiple strategies have been employed by experts to anticipate health issues before they manifest symptoms. Consequently, in the medical and healthcare communities, getting a precise diagnosis and prognosis of tumors is considered a difficult endeavour for doctors. This research thus addresses the need for medical image analysis using CAD for early diagnosis and prognosis in the healthcare domain. The literature survey communicated cutting-edge research disseminated in breast malignancy using ML and DL approaches. Although malignancy can't be proven without biopsy, early carcinoma detection using imaging modalities is an hour of need. Mammography is employed as the "benchmark" for breast carcinoma examination, owing to its widespread availability and cost-effectiveness compared to others. Current research limitations suggest that technical and practical investigation is desperately needed to boost healthcare over the long term. ‘Transnet’ is the first CAD model proposed in this research to diagnose and classify breast carcinoma with enhanced performance. Two experiments were performed on the Curated Breast Imaging Subset-Digital Database of Screening Mammography (CBIS-DDSM) dataset. The following Deep Neural Networks(DNN) were utilized VGG-16, VGG-19, Mobile Net, ResNet-50, ResNet-152, and DenseNet-169. In the first experiment, namely Deep feature fusion with ML Classifier, pre-trained networks vi were deployed as feature extractors, and afterward, the acquired attributes were provided to machine learning classifiers for classification. The second experiment, called Deep feature fusion with Neural Net classifiers, fine-tuned these networks for feature extraction and categorization. The findings revealed that the proposed approach performed remarkably well than the other cutting-edge methodologies. The second approach performed better than the first, thus, improving all the evaluation metrics. Another CAD framework proposed to enhance performance through smaller datasets is the Multi Stage Transfer Learning Approach (MSTLA). Three mammography datasets were utilized: Mammography Image Analysis Society (MIAS), In-Breast, and CBIS-DDSM. The model was fine-tuned in three stages on separate datasets, and the optimized DCNN was carried forward at the next stage. Two DNNs were deployed for training the model – DenseNet-169 and ResNet-152. The results have shown that Stage 3 performs best compared to the other two stages, with DenseNet-169 having accuracy and AUC values of 100 and 1.0. Thus, the proposed approach could be employed for early-stage breast carcinoma diagnosis. DNNs can memorize the training information owing to their huge learning capacity. In the medical domain, there is an urgent need to assess the generalizability of deep neural networks. Generalization is an approach to analyze how the model behaves on unseen data. Generalization Error(GE) measures the difference between training and testing errors. Thus to address this gap, we have proposed another framework for evaluating the generalization error in DNN. Gaussian, Salt and pepper, and Speckle noise were added to the CBIS-DDSM dataset. Generalizability was evaluated for three DCNNs - InceptionNet v3, DenseNet-201, and EfficientNet-B4. Results have shown that the proposed framework with DenseNet-201 has minimum generalization error and thus exhibits high generalizability on the unseen i.e. noisy data. This research work successfully provides a more reliable, efficient, and optimal approach for early-stage breast cancer diagnosis and thus could be deployed in laboratories. Future perspectives of the proposed methodology include its implementation on various imaging modes such as Ultrasound, MRI, CT, etc.