Optimised Convolution Layers of DnCNN using Vedic Multiplier and Hyperparameter Tuning in Cancer Detection on Field...

Area of Science:

• Medical Imaging
• Computer Science
• Artificial Intelligence

Background:

• Deep learning (DL) algorithms utilize Arithmetic Units (AU) for image processing, but fixed precision limits DL accuracy due to quantization errors.
• Cancer cell segmentation accuracy is often reduced by these quantization errors inherent in traditional DL hardware.

Purpose of the Study:

• To enhance breast cancer detection, segmentation, and classification accuracy.
• To address limitations of fixed-precision arithmetic in DL for medical image analysis.

Main Methods:

• Replaced standard multiplication with Vedic multiplication in the convolution layers of the deep convolutional neural network (DnCNN).
• Optimized the Vedic multiplication-based DnCNN architecture using the Pelican Optimization Algorithm (POA).
• Implemented the optimized POA-DnCNN on a Field-Programmable Gate Array (FPGA) for real-time breast cancer analysis.

Main Results:

• The Hybrid-Vedic (HV) multiplier, 'CUTIN,' integrated into the DnCNN's convolution layer, processes floating-point operations.
• The HV-FPGA system achieved 96.3% accuracy, 94.54% precision, 92.37% specificity, 93.56% F-score, 94.78% IoU, and 95.45% DSC.
• Outperformed existing methods in detecting breast cancer stages and classifying lesions.

Conclusions:

• The proposed CUTIN multiplier, combining Vedic mathematics principles with a carry-save adder (CSA) and simplified sum-carry generation logic (CSCGL), offers improved precision and speed.
• The optimized POA-DnCNN on FPGA provides an efficient and accurate solution for breast cancer detection.
• The system demonstrates potential for reducing area-delay and enhancing processing speed in medical imaging applications.