Table 1 Limitations in the previous studies.
From: Privacy preserving skin cancer diagnosis through federated deep learning and explainable AI
Authors | Method | XAI Used | FL | Accuracy | Limitation |
|---|---|---|---|---|---|
Srinivasu et al.18 | MobileNetV2 and LSTM | × | × | 85.34% | Reduced accuracy with noisy or high-resolution image data due to reliance on MobileNetV2 and LSTM |
Jayapriya & Jacob19 | DL method | × | × | 85.3% | 900 images as training data are too small to validate the accuracy. Further, a low accuracy of 85.3% may result inappropriate prediction |
Ding et al.21 | Lesion segmentation method | × | × | 85.1% | Small size and an Imbalanced Dataset. Further, a low accuracy of 85.1% may lead to inappropriate prediction |
Lee et al.22 | FL/conventional DL | × | ✓ | 71% 77% 66% 76% 80% | The generalizability of FL performance may remain uncertain. Use of thyroid ultrasound images further limits its applicability to other image types |
Agbley et al.23 | FL/Centralized Learning (CL) | × | ✓ | 83.01% 83.74% | Compromised performance, not used FedAvg. A lower accuracy rate may lead to an inappropriate prediction |
Gouda24 | Resnet50-Inception InceptionV3 | × | × | 84.1% 85.7% | The dataset is relatively small for training deep networks. InceptionV3 and ResNet50 require significant computational resources |
Sae-Lim25 | MobileNet | × | × | 83.9% | Used the HAM10000 dataset, which has an imbalanced class distribution Limited Convolutional layers (up to five layers of classical MobileNet) |
