Search papers, labs, and topics across Lattice.
This study evaluates five convolutional neural network (CNN) architectures, including a customized baseline and four pre-trained models, for the classification of multi-class brain tumors using a dataset of approximately 10,000 MRI images. The research found that EfficientNetB0 achieved the highest overall classification accuracy at 95%, significantly improving meningioma detection with a recall rate of 89%, compared to just 20% for simpler CNNs. Notably, the results suggest that architectural efficiency may outweigh depth in CNN performance for medical imaging tasks, as evidenced by the poorer performance of the deeper VGG19 compared to VGG16.
EfficientNetB0 outperformed other CNN architectures by achieving a remarkable 95% accuracy in brain tumor classification, highlighting the critical role of model efficiency over depth in medical imaging.
Despite recent advancements in deep learning, accurately classifying brain tumors from MRI images continues to pose challenges. In this research, we present a comprehensive evaluation of five different convolutional neural networks (CNN) architectures, including a customized baseline model and four pre-trained models - for use in classifying multi-class brain tumors using a clinically-sourced dataset of approximately 10,000 MRI images. We have utilized five different architectures; VGG16, VGG19, DenseNet121, and EfficientNetB0, which were all tested and trained within an identical experimental framework. Performance was measured by both overall accuracy and tumor-wise recall as a means to measure the clinically-relevant performance of each architecture. We found that EfficientNetB0 had the best overall classification accuracy at 95%, when compared to the other architectures tested; specifically VGG16 (94.37%), VGG19 (92.29%), DenseNet121 (90.91%) and the customized CNN (78.00%). An especially important finding of our research was the considerable improvement in detecting meningiomas; specifically, while simple CNNs could detect meningiomas with a recall rate of approximately 20%, EfficientNetB0 was able to detect meningiomas with a recall rate of 89%. Meningiomas are often difficult to detect because they can appear very subtly on MRI images. Additionally, an interesting finding was that the deeper VGG19 performed worse than the shallower VGG16. This indicates that in many cases the architectural efficiency of a CNN model may be more important than its depth when working with medical images. Overall, EfficientNetB0 appears to provide the optimal trade-off between classification accuracy, number of parameters used in the model and clinically meaningful performance.
