An Embedded Solution for Real-Time Implementation of a Deep Learning Model for Malicious Breast Tumour Detection
S. Malarvizhi, R. Kayalvizhi, H. Heartlin Maria, Revathi Venkatraman, Shatanu Patil, A. Maria Jossy · 2023
Breast cancer originates when breast cells grow and divide in an uncontrolled manner, forming a tumour. Breast cancer may not generate any symptoms in its early stages. It has been reported that early detection and prognosis are effective and are therefore crucial in breast cancer diagnosis. Different modalities are available for the screening of breast cancer, such as mammograms, ultrasound, magnetic resonance imaging (MRI), and histopathology. Every year, the number of cases seems to be increasing rapidly, resulting in advancements in various diagnostic tools and technologies. The significant increase in the global mortality rate has created an enormous opportunity to develop and implement cutting-edge computer-aided diagnostic (CAD) systems for early detection and decision-making by healthcare professionals, which is the need of the hour. The advancement of imaging technology spawned a new generation of deep learning algorithms. This study is one such artificial intelligence (AI) solution deep learning approach that uses deep learning methods to identify and categorize breast cancer from multi-modality images, namely mammogram, histopathology, magnetic resonance imaging, and ultrasound. An ensemble model comprising three different deep learning networks – namely VGG-16, ResNet-50 and Inception V3 – were trained and tested and their performance was studied. Furthermore, the developed and trained ensemble model is accelerated using PYNQ hardware accelerators for optimal diagnostic performance. The developed prototype was tested on benchmark datasets as well as real-time cases and yielded significant diagnostic results, which are discussed in this chapter. The average time taken by the ensemble model to diagnose an image is 10.4 milliseconds and the average power consumption required for this prototype is as little as 3.4 watts. This model can be used as an effective and reliable automated CAD tool for the diagnosis of breast carcinoma and can thereby aid healthcare professionals in decision-making.