An in-depth study to fine-tune the hyperparameters of pre-trained transfer learning models with state-of-the-art optimization methods: Osteoarthritis severity classification with optimized architectures

Abstract

Discrete & continuous optimization constitutes a challenging task and generally rises as an NP-hard problem. In the literature, as a derivative of this type of optimization issue, hyperparameter optimization of transfer learning (TL) architectures is not efficiently analyzed as a detailed survey in the literature. In this paper, the optimized TLbased models are effectively examined to handle this issue which constitutes the main aim of our study. For evaluation, knee osteoarthritis (KOA - a chronic degenerative joint disorder) dataset is handled to perform two challenging classification tasks which reveal the second aim of our study, i.e. binary- and multi-categorizations on KOA X-ray images. To fine-tune the hyperparameters of TL models, state-of-the-art optimization methods are chosen and compared on this competitive - NP-hard problem. Sixteen optimized architectures are designed using four efficient optimization methods (ASPSO, CDW-PSO, CSA, MSGO) and four oft-used TL models (MobileNetV2, ResNet18, ResNet50, ShuffleNet) to classify the X-ray KOA images. Regarding the experiments on both categorization tasks, the MSGO algorithm arises as more robust to be considered for hyperparameter tuning of TLbased models by achieving reliable performance. In addition, it's seen that MobileNetV2 and ResNet-based models come to the forefront for X-ray imaging-based classification by achieving high accuracy rates due to the usage of residual blocks. Consequently, in terms of mean accuracy, ResNet50-MSGO and MobileNetV2-CSA respectively record 93.15 % and 93.29 % success rates on multiclass categorization, while ResNet18-CDW-PSO and MobileNetV2-MSGO provide the same highest score of 99.43 % on binary categorization.