Author: Shanli Ding, Osama R. Mawlawi, Tinsu Pan 👨🔬
Affiliation: UT MD Anderson Cancer Center 🌍
Purpose:
Reliable detection of anomalies in Gamma Camera/SPECT flood images is vital for quality assurance (QA). Traditional methods relying on numerical thresholds and manual inspections often miss subtle anomalies. Building on prior work using the EfficientNetV2 CNN model, this study introduces contrastive pre-training within the EfficientNetV2 pipeline to optimize the anomalies detection in SPECT flood images
Methods:
Our implemented optimized approach EfficientNetV2’s classifier layer was replaced with an identity layer for feature extraction, followed by a projection head with three fully connected layers for contrastive training. A classifier head with four fully connected layers, LeakyReLU activations, batch normalization, and dropout was added for classification. The dataset included 1,020 Siemens SPECT flood images (500 normal, 500 abnormal for training; 15 abnormal, 5 normal for testing). Data augmentations, including color jittering and dimension resizing, were applied to generate pairs for contrastive learning. NT-Xent loss optimized the similarity between features. After contrastive pre-training, the backbone and projection head were fine-tuned by adding a classifier head, which was trained using cross-entropy loss. The performance of the optimized approach was compared with the regular approach using accuracy and F1-score.
Results:
The optimized and the regular approach showed an accuracy and F1-score: 0.94/0.92 and 0.95/0.87, respectively, demonstrating improved performance with the contrastive pre-training stage.
Conclusion:
By leveraging contrastive pre-training, the fine-tuned approach improves accuracy and detects subtle anomalies, addressing limitations of the regular method. This advancement enhances nuclear medicine QA workflows through reliable, automated anomaly detection. Future work will focus on expanding the dataset and refining the framework to improve robustness and generalizability.