DeiT Outperforms Experts in Cancer Diagnosis
Created on January 2|Last edited on January 2
Comment
A recent study in Nature Medicine showcased the application of advanced transformer-based neural networks in diagnosing ovarian cancer. Using the DeiT (Data-efficient Image Transformer) architecture, the models demonstrated remarkable performance in distinguishing between benign and malignant ovarian lesions, surpassing human examiners and offering robust generalization across diverse clinical settings.
Model Training and Architecture
The study utilized the DeiT-S (small) architecture, which was pretrained on ImageNet for transfer learning. This approach, widely adopted in medical imaging, significantly enhanced the model’s capacity to interpret ultrasound images. The model architecture featured an output layer with ten nodes corresponding to specific histological categories within benign and malignant classes, allowing it to leverage detailed diagnostic information.
Training was conducted on a dataset comprising 17,119 ultrasound images from 3,652 patients. A leave-one-center-out cross-validation scheme was employed, ensuring robust generalization by training the model on data from all but one center and testing it on the excluded center. Within each iteration, the remaining data were split into training (90%) and validation (10%) sets, balanced to include equal numbers of benign and malignant cases in the validation set.
The models were trained using multiclass focal loss, which emphasized misclassified examples more effectively than traditional cross-entropy loss. The AdamW optimizer and with a progressively reduced learning rate when validation performance plateaued.
Input images were preprocessed to enhance model training efficiency. Images were cropped to regions of interest, resized to 256×256×3 pixels, and normalized using precomputed pixel statistics. During training, random cropping reduced the images further to 224×224×3 pixels. Data augmentation via RandAugment introduced five random transformations per image, excluding color transformations.
Model Performance
The AI models achieved an F1 score of 83.5%, significantly higher than the scores of expert (79.5%) and non-expert (74.1%) examiners. Sensitivity and specificity reached 89.31% and 88.83%, respectively, both surpassing human performance. These results reflect the ability of the models to minimize false negative and false positive rates, crucial for accurate diagnoses.
Conclusion
The integration of transformer-based models like DeiT-S into clinical workflows marks a significant advancement in medical imaging. By leveraging transfer learning and robust preprocessing techniques, these models have set a new benchmark for ovarian cancer diagnosis. Further prospective studies are anticipated to validate their clinical applicability and expand their use in other diagnostic domains.
Add a comment
Tags: ML News
Iterate on AI agents and models faster. Try Weights & Biases today.