Optimised Convolution Layers of DnCNN Using Vedic Multiplier and Hyperparameter Tuning in Cancer Detection on Field Programmable Gate Array.

Introduction: Recently, deep learning (DL) algorithms use Arithmetic Units (AU) in CPU/GPU hardware for processing images/data. AU operates in fixed precision and limits the representation of weights and activations in DL. The problem leads to quantization errors, which reduce accuracy during cancer cell segmentation.

Methods: In this study, arithmetic multiplication in convolution layers is replaced with Vedic multiplication in the proposed DnCNN algorithm. Next, Vedic multiplication-based convolution layers in the DnCNN architecture are optimized using POA (Pelican Optimization Algorithm), and the resulting POA-DnCNN is implemented on an FPGA device for breast cancer detection, segmentation, and classification of benign and malignant breast lesions.

Discussion: In the convolution layer of DnCNN, floating-point operations are performed through the Hybrid-Vedic (HV) multiplier called 'CUTIN,' which is the combination of Urdhva Tryambakam and Nikhilam Sutra with the upasutra 'Anurupyena.' Larger image sizes increase processor size and gate count.

Results: The proposed HV-FPGA-based breast cancer detection system, employing Vedic multiplication in the convolution layers of DnCNN and hyperparameters optimized by POA, detects stages of breast cancer with an accuracy of 96.3%, precision of 94.54%, specificity of 92.37%, F-score of 93.56%, IoU of 94.78%, and DSC of 95.45%, outperforming existing methods.

Conclusion: The proposed CUTIN multiplier uses a CSA (carry save adder) with simplified sum-carry generation logic (CSCGL), achieving lower area-delay, high speed, and improved precision.