Interactive Machine Learning-Based Multi-Label Segmentation of Solid Tumors and Organs.

Dimitrios Bounias,Russell T Shinohara,Benjamin A Greenberger,Joseph Lombardo,Nariman Jahani,Aimilia Gastounioti,Michelle Hershman,Leonid Roshkovan,Sarthak Pati,Sharyn I Katz,Hamed Akbari,Michel Bilello,Christos Davatzikos,Saima Rathore,Despina Kontos,Spyridon Bakas,Bardia Yousefi,Amber L Simpson,Ashish Singh,Konstantina S Nikita ,Lou C ,Rhea Chitalia ,K G Richard

doi:10.3390/app11167488

Abstract

We seek the development and evaluation of a fast, accurate, and consistent method for general-purpose segmentation, based on interactive machine learning (IML). To validate our method, we identified retrospective cohorts of 20 brain, 50 breast, and 50 lung cancer patients, as well as 20 spleen scans, with corresponding ground truth annotations. Utilizing very brief user training annotations and the adaptive geodesic distance transform, an ensemble of SVMs is trained, providing a patient-specific model applied to the whole image. Two experts segmented each cohort twice with our method and twice manually. The IML method was faster than manual annotation by 53.1% on average. We found significant (p < 0.001) overlap difference for spleen (DiceIML/DiceManual = 0.91/0.87), breast tumors (DiceIML/DiceManual = 0.84/0.82), and lung nodules (DiceIML/DiceManual = 0.78/0.83). For intra-rater consistency, a significant (p = 0.003) difference was found for spleen (DiceIML/DiceManual = 0.91/0.89). For inter-rater consistency, significant (p < 0.045) differences were found for spleen (DiceIML/DiceManual = 0.91/0.87), breast (DiceIML/DiceManual = 0.86/0.81), lung (DiceIML/DiceManual = 0.85/0.89), the non-enhancing (DiceIML/DiceManual = 0.79/0.67) and the enhancing (DiceIML/DiceManual = 0.79/0.84) brain tumor sub-regions, which, in aggregation, favored our method. Quantitative evaluation for speed, spatial overlap, and consistency, reveals the benefits of our proposed method when compared with manual annotation, for several clinically relevant problems. We publicly release our implementation through CaPTk (Cancer Imaging Phenomics Toolkit) and as an MITK plugin.

Highlights

Medical image segmentation is an important task in clinical and research environments [1,2,3,4], facilitating subsequent computational analyses, which depend on the accuracy of the segmentation [5,6]
We propose an Interactive machine learning (IML) method leveraging adaptive geodesic distance (AGD) [17] maps alongside an ensemble of support vector machines (SVMs) that is agnostic to image type/dimensionality
Manual and IML-assisted segmentations yielded similar pairs of Dice Similarity Coefficient (DSC) both for whole tumor (WT) and individual sub-regions, thereby indicating no significant difference between them, whereas the converse was true for other cohorts

Summary

Introduction

Medical image segmentation is an important task in clinical and research environments [1,2,3,4], facilitating subsequent computational analyses, which depend on the accuracy of the segmentation [5,6]. A plethora of fully automatic machine learning (ML) methods that can achieve stateof-the-art results have been proposed, but tend to face various challenges [7] that hinder clinical translation. Some of the most important challenges are generalization to unseen datasets and need for extensive expert corrections and refinements [4,8]. Interactive machine learning (IML) methods fill the void between manual and automatic approaches by allowing an operator to train a patient-specific model via quick and rough drawings, which automatically segments the entire scan [9,10,11]. IML approaches provide the option for expedited refinements, and the final segmentation tends to get closer to the desired result as a function of the invested time

Objectives

Results

Conclusion