Boot-DLMIR: New Methods for he Fast Development of Deep Learning-Based Radiology Solutions with Very Few Annotated Datasets

Leo Joskowicz ¹ Adi Szeskin ¹ Shalom Rochman ¹ Bella Spektor Fadida ¹ Netanell Avisdris ¹ Richard Lederman ² Jacob Sosna ² Liat Ben Sira ^3,5 Dafna Ben Bashat ^3,4

¹School of Computer Science and Engineering, The Hebrew University of Jerusalem, Israel
²Department of Radiology, Hadassah University Hospital, Israel
³Sackler Faculty of Medicine and Sagol School of Neuroscience, Tel Aviv University, Israel
⁴Sagol Brain Institute, Tel Aviv Sourasky Medical Center, Israel
⁵Division of Pediatric Radiology, Tel Aviv Sourasky Medical Center, Israel

Background: Deep learning methods have proven effective for the detection and segmentation of anatomical structures and pathologies in CT and MRI. However, ttraining of the deep network classifiers requires manual annotation of many cases by expert radiologists, which is tedious, time-consuming and impractical for most structures.

Methods: We have developed Boot-DLMIR, a set of new methods for the fast development of deep learning-based solutions in Radiology with very few annotated datasets. The key idea is to bootstrap the creation of expert-validated annotations with new techniques for annotation uncertainty estimation and for learning how experts correct annotations generated by deep learning networks initially trained with very few annotated datasets. Our methods optimize the time required by radiologists to annotate, validate, and correct segmentations by providing methods to: 1) select the order in which annotated slices are examined and corrected; 2) select which scans to annotate and validate to increase the robustness and coverage of the network based on the segmentation uncertainty estimation; 3) decrease data overfitting to one radiologist and a few annotated datasets; 4) increase the radiologist confidence in the results by providing an estimated uncertainty measure; 5) support active learning with annotated datasets of mixed quality. We evaluate our method on automatic detection and segmentation of fetal brain, fetal body and placenta on MRI and of liver and liver tumors on CECT with the Dice coefficient.

Results: Our methods achieved a Dice coefficient whose value is above the manual observer annotation variability (number of annotated scans used for training): 0.997±0.03 for the fetal body (45), 0.966±0.03 for the fetal brain (4530), 0.843±0.10 for the placenta (16), 0.95±0.03 for the liver (109) and 0.82±0.14 for liver tumors (126).

Conclusion: Boot-DLMIR may accelerate the development of accurate and reliable deep learning based methods in Radiology using very few annotated datasets.