A Framework for Generalization Error Evaluation in Deep Convolutional Neural Networks
Gunjan Chugh, Shailender Kumar, Nanhay Singh · 2023
Deep learning has introduced various paradigms in the healthcare industry. Deep Convolutional Neural Network models assist doctors in diagnosis, surgery, and other areas. Incidences of breast tumors are growing at a frightening pace. Thus, it is necessary to diagnose it so that the mortality count can be decreased. Generalizability defines how well the model performs on unseen data. When capturing mammograms different types of noise get added to the images. Noise may significantly diminish classification ability and make class separation more difficult. Thus, analyzing the model's generalization on noisy or unseen data is very important. This paper proposes an approach for analyzing generalization errors in Deep Convolutional Neural Networks by inducting noises such as Gaussian, Salt and pepper, and Speckle. We have utilized the CBIS-DDSM dataset. Three prominent deep neural network models- Inception v3, Dense Net 201, and EfficientNetB4- are fine-tuned to assess the model's efficiency on noisy data. Results proved that the DenseNet201 has minimum Generalization errors i.e. 0.12 and performs pretty well on noisy data with a minimum loss rate. On the other hand, maximum distortion is caused by Speckle Noise in Inceptionv3 and Efficient NetB4 leading to a considerable drop in accuracy. Inception v3 has the highest Generalization Error i.e. 0.61 and thus exhibits minimum generalization capability.