Centrala begrepp
The author explores different methods for out-of-distribution detection in breast cancer classification using point-of-care ultrasound imaging, emphasizing the importance of reliable assessments and safe classifiers.
The main thesis is to compare and evaluate three OOD detection methods - softmax, energy score, and deep ensembles - to enhance the accuracy of breast cancer classification in POCUS images.
Sammanfattning
In the study, various OOD detection methods were compared to improve breast cancer classification accuracy using point-of-care ultrasound imaging. The research focused on detecting unreliable assessments through softmax, energy score, and deep ensemble methods. Results showed that the ensemble method was the most robust across all OOD data sets. The study highlighted the significance of balancing performance and computational complexity in OOD detection for real-world medical applications.
The ID data comprised POCUS images of normal tissue, benign, and malignant lesions collected at Sk˚ane University Hospital. Three different OOD test data sets were used: MNIST, CorruptPOCUS, and CCA. The study implemented a CNN architecture with five convolutional layers for classification purposes. Metrics such as ROC curves, AUC, FPR were used to evaluate the performance of each method. The results indicated that the ensemble method outperformed softmax and energy score methods across all OOD data sets.
The study concluded that while softmax and energy score methods are easier to implement without retraining networks, ensembles offer more robust results but require higher computational power. Further research on Bayesian neural networks and other OOD detection methods is recommended for future investigations.
Statistik
Train (POCUS)
Train (US)
Test (POCUS)
Normal: 304 / 168 / 284
Benign: 140 / 101 / 131
Malignant: 125 / 398 / 116
Total: 569 / 667 / 531
Citat
"The ensemble method is the most robust, performing the best at detecting OOD samples for all three OOD data sets."
"Softmax probabilities have been used for OOD detection based on low scores for predicted classes."
"Energy score method outperforms softmax method on two of the data sets."