A deep learning algorithm for automatic 3D segmentation and quantification of hamstrings musculotendon injury from MRI

Figures

• 

  Fig. 1. Segmentation atlases of the regions of interest for the HTEN (top) and EDEMA (bottom) segmentations. Shown are examples of the segmentation of the edema in different cases. Abbreviations key: ST) semitendinosus, SM) semimembranosus, BF) biceps femoris, LH) long head, SH) short head, JT) joint tendon, PT) proximal tendon, DT) distal tendon.

• 

  Fig. 2. Visual representation of the methods. (A) Color-coded three-dimensional representation of the HTEN regions of interest (ROIs) and EDEMA ROI segmented, trained by the AI, and used for subsequent analysis. (B) Workflow of the HTEN and EDEMA AI models.

• 

  Fig. 3. Visual representation of the different metrics quantified by the segmentations derived by (A) EDEMA scans (left), (B) HTEN scans (middle), and (C) the combination of EDEMA and HTEN overlayed (right).

• 

  Fig. 4. Example investigation of AI output compared to the ground-truth for the volume metric comparison correlation (left) and Bland-Altman analysis (right) for (A) the EDEMA analysis edema ROI and (B) an example tendon ROI (Biceps Femoris Semitendinosus tendon) from the HTEN analysis. Note: In Fig. 4B, the biceps femoris-semitendinosus joint tendon exhibits several outlying points due to morphological variability at common injury sites. In Fig. 4A (right), the high-volume outlier (~ 250 ml) corresponds to a large athlete with extensive edema visible on imaging.

• 

  Fig. 5. (A) Relating the AI label outputs to BAMIC grade for (left) total edema label volume and (right) max edema content (percentage of edema volume in total muscle volume, %) found in a muscle for A & B sites (myofascial and myotendinous junction/muscular sites) and C sites (tendinous sites) with an unpaired t-test. Color-mapping of maximum muscle impact (right) indicates the volume of edema overlapping tendon pixels. (B) Repeat of (A) however utilizing the manually segmented labels for the calculation instead of the AI labels. (C) Sample hamstring tendon and edema segmentations (color coding shown) categorized by BAMIC site demonstrating various presentations of the three sites. Full results relating other edema AI segmentation output metrics to BAMIC grades are in supplementary material.