In this preliminary study,
we aimed to train a deep learning model to detect the presence of pelvic fracture or dislocation on routine trauma antero-posterior radiographs.
This initial proof of concept demonstrated some early network trainability given the limited training gold standard of single reviewer labels.
Generalizability of this system could perhaps be improved with a more comprehensive dataset and and consensus review of training labels.
Early AUC results are encouraging.
However,
even if an AUC of 0.888 is similar to prior bone fracture studies (Olczac,
2017; Kim 2017; Gale,
2017; Urakawa,
2018; Adams,
2018),
we remain far from outperforming human-expertise and believe that the current level of detection and localization performance falls far short of the ideal for clinical workflow prioritization.
This impression is derived not so much from objective AUC analysis,
but from the subjective heat-map visualization impressions of fractures both detected and missed by the current system.
We also noted that when analysing class activation maps that it was possible for a network to report a correct classification,
but generate discordant heatmaps with no reliable identification of a fracture or dislocation.
Of the various forms of fractures,
the most successful examples concerned large displaced intertrochanteric femoral fractures,
with significantly greater difficulty with pelvic,
acetabular,
and diathesis disruptions.
While some of this may be attributable to the lower frequency of these other forms of fractures,
we also feel there may also be something fundamental about the visual manifestations of these findings that pose a challenge to naive detection and localization with convolutional neural networks,
particularly where the findings are radiolucent.
Among these differences,
the pubic symphysis dislocation for instance appeared the most difficult injury to detect,
highlighting the limits of a quadrant approach for pelvic radiographs,
and in some ways paralleling the difficulty of subtle findings for even human-expert readers.
Nonetheless,
these results do raise interesting possibilities for automated plain film triage workflows.
From the subjective review of generated visualization heat-maps,
it was felt that localization success was greatest for the most frequent and highly displaced large fracture fragments,
an impressive localization effort,
given the relatively few number of training images,
and lack of predefined training bounding boxes provided for the neural network.
It was also felt that the limited sensitivity could be related to a constantly varying patient positioning of what would be considered normal,
for instance,
the changing position of the coxo-femoral joint between the quadrants.
which would impede the training of a generalizable neural network.
While some of this variance could have been addressed with an exclusion of low-quality or poorly positioned radiographs,
it was our intent with this investigation to investigate network performance given more undifferentiated clinically relevant presentations,
encountered in any chaotic resuscitation or trauma bay and examine performance in that more realistic setting.
Further Investigations:
Based upon these preliminary results,
we believe that further optimization to both the labeling data or the utilization of multi-scale pyramidal neural network technologies such as RetinaNet (2017) could improve network localization and sensitivity.
In addition,
given the variable performance with radiolucent findings,
a far more intensive but rigorous approach involving the segmentation and contouring of normal osseous structures and boundaries may more closely mimic the analytic process of a human expert interpreter and be more suitable for certain more challenging subclasses of traumatic injuries and findings.
Finally,
an automated pre-processing process taking into variant patient positioning and the ability to form a weak segmentation or co-registration to normal anatomic views may further improve classification and localization performance.
Conclusion:
In conclusion,
these preliminary results showed that DCNN transfer learning approach is able to some extent identify automatically identify,
localize and triage cases of acute fracture or dislocation on pelvic radiographs.
Further investigation is required to further elucidate and address other more challenging traumatic injury manifestations to further improve network performance for routine clinical application.