Table 3.

A comparison of the strengths and weaknesses of the suggested and previous methods.

Author	Strengths	Weaknesses
Kufel et al.	Highest average score among similar research; use of the custom data split with a random seed; EfficientNet used as a backbone model.	Use of EfficentNet B1 version due to GPU limitations (two Nvidia 1080 Ti).
Yao et al. [12]	Use of CNN capsules; the best ROC AUC in infiltration and pneumonia achieved.	Challenges in effective training with small datasets, and limited applicability compared to traditional neural network architectures.
Shen et al. [14]	Better results in classification of infiltration and pneumonia (compared to Kufel et al.).	Method includes its complexity in terms of computational requirements due to the multiple layers, as well as the potential challenges in efficiently updating coupling coefficients and performing dynamic routing in the context of 1 × 1 convolutional layers, which may lead to computational exhaustiveness and limitations in feasibility for practical implementations.
Yan et al. [24]	Better results in classification of emphysema, fibrosis, and nodule (compared to Kufel et al.).	Approach involves a reliance on comparisons with other methods, where methodological variations such as data splitting setups and additional disease information can impact the fairness and interpretability of the performance assessment. Moreover, while the method demonstrates improvements in overall performance and addresses challenges like spatial squeezing and lesion size variation, it might not fully address nuanced differences in disease presentation and diagnostic intricacies present in real-world scenarios.
Güendel et al. [19]	Use of location-aware DNN, combining spatial information and high-resolution images.	Method includes a potential susceptibility to overfitting due to the use of a complex architecture and high-resolution images, especially given the imbalance in the dataset. Additionally, the approach might face challenges in accurately representing complex spatial information and disease locations, particularly when dealing with multiple and diffuse diseases for which precise position information is lacking.
Wang et al. [20]	Larger GPU was used for the training process.	Worse results achieved in each disease and worse mean AUC ROC (compared to Kufel et al.).
Baltruschat et al. [25]	Larger GPU was used in the training process, better results in classification of hernia (compared to Kufel et al.).	The weaknesses associated with this method include potential challenges arising from the use of transfer learning and fine-tuning. While transferring knowledge from a different domain might provide a head start, it could also introduce biases or assumptions from the source domain that do not hold true in the medical context, potentially leading to suboptimal performance or misinterpretations. Furthermore, adapting complex architectures like ResNet-50 for medical imaging with limited datasets may increase the risk of overfitting, as the model might be highly sensitive to variations in the small dataset. Additionally, although incorporating patient-specific information (age, gender, view position) may seem advantageous, it could also introduce noise or irrelevant factors that may not consistently aid in improving classification performance across various scenarios.