Abstract
Purpose :
Deep learning enhanced computer aided diagnoses have the potential to help increase the efficiency of image heavy clinical workflows. However, performance of medial deep learning is often limited by (i) lack of large medical training sets due to the high cost of annotations (ii) the heterogeneous, and subtle feature differences in objects of interest. In this work we propose that the inherently structured anatomical topology of a given imaging modality can be leveraged to allow a CNN to learn a more robust embedding by contrasting pairs of normal and abnormal images thereby improving the performance of object detection on pathological features.
Methods :
This project leveraged open source Heidelberg Spectralis AMD OCT*, with 600 images expertly annotated for 9 classes of features of interest were. A YOLOv4 network was adapted to contain a novel contrastive induced gated attention module (CIGA) which leverages the anatomical topology to aid detection of pathological objects. The baseline YOLOv4 network was trained on the OCT images and the performance of the CIGA module was assessed by fine-tuning the baseline with the CIGA. The training process was performed with multiple repeats.
*Reza et al, IEEE Transactions on Medical Imaging, 37(4):1024-1034, (2018)
Results :
The fine-tuning with the CIGA module produced better overall MAP50 values than the baseline network alone. The CIGA network produced an overall test MAP50 of 0.654(0.023) compared to the baseline performance of 0.635(0.01) and a test mAP50:95 of 0.369(0.007) compared to the baseline performance of 0.355(0.01). Performance was also assessed for the nine classes of pathological feature such as Hyperfluorescent spots where the CIGA achieved an MAP50 of 0.732(0.049) compared to the baseline 0.680(0.021) and PR layer disruption where the CIGA achieved an MAP50 of 0.712(0.095) compared to the baseline performance of 0.635(0.018).
Conclusions :
In this work we propose the CIGA network and demonstrate incremental performance gains for the task of object detection in OCT scans. This was achieved by leveraging the structured topological information in the anatomy with the help of gated attention under a contrastive learning framework to enhance the signal of the different pathological objects. This method will improve detection of small biomarkers from imaging, which are traditionally difficult to detect using traditional deep learning methods.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.