PV-0531: Multi-centre evaluation of atlas-based and deep learning contouring using a modified Turing Test

Clinical evaluation of atlas-based auto-segmentation in breast and nodal radiotherapy

OBJECTIVES

Accurate contouring of anatomical structures allows for high-precision radiotherapy planning, targeting the dose at treatment volumes and avoiding organs at risk. Manual contouring is time-consuming with significant user variability, whereas auto-segmentation (AS) has proven efficiency benefits but requires editing before treatment planning. This study investigated whether atlas-based AS (ABAS) accuracy improves with template atlas group size and character-specific atlas and test case selection.

METHODS AND MATERIALS

One clinician retrospectively contoured the breast, nodes, lung, heart, and brachial plexus on 100 CT scans, adhering to peer-reviewed guidelines. Atlases were clustered in group sizes, treatment positions, chest wall separations, and ASs created with Mirada software. The similarity of ASs compared to reference contours was described by the Jaccard similarity coefficient (JSC) and centroid distance variance (CDV).

RESULTS

Across group sizes, for all structures combined, the mean JSC was 0.6 (SD 0.3, p = .999). Across atlas-specific groups, 0.6 (SD 0.3, p = 1.000). The correlation between JSC and structure volume was weak in both scenarios (adjusted R2-0.007 and 0.185).Mean CDV was similar across groups but varied up to 1.2 cm for specific structures.

CONCLUSIONS

Character-specific atlas groups and test case selection did not improve accuracy outcomes. High-quality ASs were obtained from groups containing as few as ten atlases, subsequently simplifying the application of ABAS. CDV measures indicating auto-segmentation variations on the x, y, and z axes can be utilised to decide on the clinical relevance of variations and reduce AS editing.

ADVANCES IN KNOWLEDGE

High-quality ABASs can be obtained from as few as ten template atlases.Atlas and test case selection do not improve AS accuracy.Unlike well-known quantitative similarity indices, volume displacement metrics provide information on the location of segmentation variations, helping assessment of the clinical relevance of variations and reducing clinician editing. Volume displacement metrics combined with the qualitative measure of clinician assessment could reduce user variability.

Beyond automatic medical image segmentation-the spectrum between fully manual and fully automatic delineation

Semi-automatic and fully automatic contouring tools have emerged as an alternative to fully manual segmentation to reduce time spent contouring and to increase contour quality and consistency. Particularly, fully automatic segmentation has seen exceptional improvements through the use of deep learning in recent years. These fully automatic methods may not require user interactions, but the resulting contours are often not suitable to be used in clinical practice without a review by the clinician. Furthermore, they need large amounts of labelled data to be available for training. This review presents alternatives to manual or fully automatic segmentation methods along the spectrum of variable user interactivity and data availability. The challenge lies to determine how much user interaction is necessary and how this user interaction can be used most effectively. While deep learning is already widely used for fully automatic tools, interactive methods are just at the starting point to be transformed by it. Interaction between clinician and machine, via artificial intelligence, can go both ways and this review will present the avenues that are being pursued to improve medical image segmentation.

Interactive contouring through contextual deep learning

PURPOSE

To investigate a deep learning approach that enables three-dimensional (3D) segmentation of an arbitrary structure of interest given a user provided two-dimensional (2D) contour for context. Such an approach could decrease delineation times and improve contouring consistency, particularly for anatomical structures for which no automatic segmentation tools exist.

METHODS

A series of deep learning segmentation models using a Recurrent Residual U-Net with attention gates was trained with a successively expanding training set. Contextual information was provided to the models, using a previously contoured slice as an input, in addition to the slice to be contoured. In total, 6 models were developed, and 19 different anatomical structures were used for training and testing. Each of the models was evaluated for all 19 structures, even if they were excluded from the training set, in order to assess the model's ability to segment unseen structures of interest. Each model's performance was evaluated using the Dice similarity coefficient (DSC), Hausdorff distance, and relative added path length (APL).

RESULTS

The segmentation performance for seen and unseen structures improved when the training set was expanded by addition of structures previously excluded from the training set. A model trained exclusively on heart structures achieved a DSC of 0.33, HD of 44 mm, and relative APL of 0.85 when segmenting the spleen, whereas a model trained on a diverse set of structures, but still excluding the spleen, achieved a DSC of 0.80, HD of 13 mm, and relative APL of 0.35. Iterative prediction performed better compared to direct prediction when considering unseen structures.

CONCLUSIONS

Training a contextual deep learning model on a diverse set of structures increases the segmentation performance for the structures in the training set, but importantly enables the model to generalize and make predictions even for unseen structures that were not represented in the training set. This shows that user-provided context can be incorporated into deep learning contouring to facilitate semi-automatic segmentation of CT images for any given structure. Such an approach can enable faster de-novo contouring in clinical practice.