Abstract
Inspired by the hierarchical, receptive-field-based processing of biological vision and by synaptic pruning in the developing brain, this work proposes a bio-inspired optimized convolutional neural network framework for multi-class cancer classification from RNA-Seq gene expression data. Although current deep learning models achieve strong predictive performance, existing approaches typically rely either on large pre-trained networks that cannot be deployed on resource-constrained devices or on image representations that normalize gene expression values to the 8-bit range (0–255), discarding the original biological dynamic range. A variance-driven feature selection step reduces 20,531 gene profiles to 16,384 high-information features, which are then reshaped into 128×128 images while preserving the original log2-normalized expression range (approximately 0–20.78) without rescaling to that 8-bit range. The network architecture is derived from a parametric mathematical model that enables analytical control over parameter count and FLOPs, resulting in a compact network with only 0.16 million parameters, a model size of 1.29 MB, and 0.0372 GFLOPs. On the TCGA PANCAN dataset (five cancer classes), the proposed baseline achieves 99.38% accuracy, matching or exceeding twelve state-of-the-art pre-trained architectures. For real-world deployment, the network was further optimized through two pathways: (A) direct TensorFlow Lite quantization in FP32, FP16, and INT8, and (B) structured channel pruning, analogous to synaptic pruning, followed by TFLite conversion. All variants retained the baseline 99.38% accuracy, with the Pruned TFLite FP16 variant reducing computational cost by 23.4% to 0.0285 GFLOPs, model size by 80.6%, and latency by 75.4%, while achieving the highest throughput of 1256 images/s at only 0.25 MB.