Automatic segmentation of head and neck primary tumors on MRI using a multi-view CNN

Abstract

Background Accurate segmentation of head and neck squamous cell cancer (HNSCC) is important for radiotherapy treatment planning. Manual segmentation of these tumors is time-consuming and vulnerable to inconsistencies between experts, especially in the complex head and neck region. The aim of this study is to introduce and evaluate an automatic segmentation pipeline for HNSCC using a multi-view CNN (MV-CNN). Methods The dataset included 220 patients with primary HNSCC and availability of T1-weighted, STIR and optionally contrast-enhanced T1-weighted MR images together with a manual reference segmentation of the primary tumor by an expert. A T1-weighted standard space of the head and neck region was created to register all MRI sequences to. An MV-CNN was trained with these three MRI sequences and evaluated in terms of volumetric and spatial performance in a cross-validation by measuring intra-class correlation (ICC) and dice similarity score (DSC), respectively. Results The average manual segmented primary tumor volume was 11.8±6.70 cm3 with a median [IQR] of 13.9 [3.22-15.9] cm3. The tumor volume measured by MV-CNN was 22.8±21.1 cm3 with a median [IQR] of 16.0 [8.24-31.1] cm3. Compared to the manual segmentations, the MV-CNN scored an average ICC of 0.64±0.06 and a DSC of 0.49±0.19. Improved segmentation performance was observed with increasing primary tumor volume: the smallest tumor volume group (<3 cm3) scored a DSC of 0.26±0.16 and the largest group (>15 cm3) a DSC of 0.63±0.11 (p<0.001). The automated segmentation tended to overestimate compared to the manual reference, both around the actual primary tumor and in false positively classified healthy structures and pathologically enlarged lymph nodes. Conclusion An automatic segmentation pipeline was evaluated for primary HNSCC on MRI. The MV-CNN produced reasonable segmentation results, especially on large tumors, but overestimation decreased overall performance. In further research, the focus should be on decreasing false positives and make it valuable in treatment planning.

Strategies for tackling the class imbalance problem of oropharyngeal primary tumor segmentation on magnetic resonance imaging

“…Their work focused on studying the value of multiparametric MRI on the segmentation performance, both for qualitative and quantitative imaging. Other works focused on the automatic segmentation on multiparametric MRI of the head and neck cancer in general, rather than on the particular subset of oropharyngeal cancer: Bielak et al [28] used diffusion weighted imaging while Schouten et al [29] proposed a multiview CNN architecture. To the best of our knowledge, only our work is focused on tackling the class imbalance problem for head and neck cancer segmentation on MRI, and particularly for the oropharyngeal subsite.…”

Magnetic Resonance Guided Radiotherapy for Head and Neck Cancers

“…One hindrance of on-table adaptation is the need to quickly delineate, or outline, the critical nearby organs at risk in a manner that is similar to physician-approved delineation while the patient is waiting to receive RT. Currently available software and approaches are based on CT images or higher-field MRI (non-MRgRT) [ 33 ]. There are no currently available autosegmentation approaches that accurately delineate important organs and targets in the head/neck region with the lower field MRI available for MRgRT, and further development of these approaches would greatly facilitate on-table adaptation.…”

Deep Learning for Automated Elective Lymph Node Level Segmentation for Head and Neck Cancer Radiotherapy

“…First, since UNet is a widely established CNN that is used for a variety of imaging-related problems [ 12 ] and since it was used in two other studies for combined lymph structure segmentation [ 9 , 13 ], we included a patch-based UNet variant as a baseline model configuration. Other works have suggested the use of voxel-classification methods for individual LN level segmentation using a 3D multi-scale network [ 14 ], as well as 2.5D (multi-view; MV) networks for several segmentation challenges (multiple sclerosis [ 15 ], ocular structures [ 16 ], abdominal lymph structures [ 17 ], head-and-neck tumors [ 18 ]). Because 2.5D networks may more effectively learn features in the presence of little data [ 19 ] and because voxel classification may better resolve local ambiguities near level transitions, a multi-view convolutional neural network (MV-CNN) was included as our second configuration.…”

“…Because 2.5D networks may more effectively learn features in the presence of little data [ 19 ] and because voxel classification may better resolve local ambiguities near level transitions, a multi-view convolutional neural network (MV-CNN) was included as our second configuration. This method, however, appears limited by a systematic over-estimation of foreground classes [ 18 ]. Therefore, as our third configuration, UNet was used for foreground segmentation, and subsequently MV was used for classifying the foreground voxels into individual LN levels.…”