Deep Learning for Cancer Detection: Efficiency of EfficientNetB3 vs CNN in Histopathological Image Classification

Exact cancer detection combined with prompt pathological image sorting serves as a requirement for both treatment procedures and disease prognosticating. The research evaluates EfficientNetB3 as a current top deep learning model against a built-from-scratch convolutional neural network (CNN) for tissue samples classification tasks. A total of 10,000 balanced histopathological images were distributed for training at 70% and validation at 20% before testing on the remaining 10%. Experimental research demonstrates that EfficientNetB3 delivers better results than other deep learning models through its precision of 96.2% while achieving identical accuracy and F1-score of 95.8% along with recall of 95.5%. The custom CNN reaches slightly lower outcomes than EfficientNetB3 with 91.3% accuracy and 91.8% precision and 90.9% recall and a 91.3% F1-score. EfficientNetB3 surpasses the discriminative power of the custom CNN because it exhibits an AUC value of 0.96 whereas the custom CNN only achieves 0.91. The custom CNN demonstrates superior processing speed compared to EfficientNetB3 by executing at 6.8 ms while EfficientNetB3 requires 12.5 ms thus making it ideal for applications with real-time and resource constraints. The confusion matrices indicate EfficientNetB3 correctly detects more than 90% of the samples from each class and exhibits lesser misidentification errors than the custom CNN. These findings are supported by visual evaluation through ROC curves as well as radial performance plots and scatter probability distributions. The study presents a relationship between precision and system performance and recommends EfficientNetB3 for critical medical applications and a tailored CNN solution for immediate practical applications that require efficient resource usage. Future research aims to combine these two architectural approaches for achieving better results between accuracy rates and inference speeds.