An alternative classification system for categorizing pediatric femoral neck fractures (PFNFs) has been shown to have significant clinical values in guiding treatment. This system, which is based on the direction of translation and the presence of posteromedial column comminution, is an alternative to the classic Delbet-Colonna system. However, this system relies on the user's clinical experience.

Due to the rarity of PFNFs, can a deep learning (DL) model trained based on limited samples accurately classify PFNFs types according to the novel classification thereby minimizing the dependence on experience?

Despite the scarcity of PFNFs, the multimodal DL model trained on the radiograph-text paired dataset can still accurately categorize PFNFs types according to the novel classification and outperform the unimodal DL model solely trained on radiographic dataset as well as the experienced pediatric orthopedic surgeons (POS).

This multicenter study retrospectively reviewed 711 anteroposterior (AP) and 674 lateral X-rays from 711 patients (mean age: 10.7 ± 3.7 years) with PFNFs treated surgically. AP X-rays were divided into 2 types (576 PFNFs without and 135 with medial comminutions), and lateral X-rays were classified into 4 types (132 Type I, 329 Type II, 91 Type III PFNFs, and 122 PFNFs with comminuted posterior columns).

X-rays-text paired datasets were developed to include each subtype and its corresponding textual description. The datasets were randomly divided into the training and the testing datasets in an approximately 4:1 ratio. The multimodal and unimodal DL models were trained based on the X-rays-text paired datasets and X-rays, respectively.

Four junior-level, 4 mid-level and 4 senior-level POS were asked to assess the X-ray subtypes of the testing datasets. The rating performance of multimodal, unimodal DL models and POS with various experience were evaluated.

When classifying the AP X-ray subtypes in the testing dataset, the multimodal DL model outperformed the unimodal DL model in terms of accuracy (96.5% versus 90.1 %), recall (92.4% versus 85.8%), positive predictive value (96.3% versus 84.0%), negative predictive value (96.3% versus 84.0%), F1 score (94.2% versus 84.8%), and the area under the curve (95.2% versus 91.6%). Similar results were found when comparing the performance of the two models in classifying the lateral X-ray subtypes of the testing dataset.

According to the testing dataset, one-way analysis of variance results indicated that senior-level POS show significantly superior accuracy in classifying AP (p < 0.001) or lateral (p < 0.001) X-rays subtypes to that of mid-level or junior-level POS. One sample t-tests results showed that the multimodal DL model's accuracy of AP or lateral X-rays typing were significantly higher than that of senior-level POS (p < 0.05).

Multimodal DL model outperforms experienced POS in PFNFs typing, reducing the need for clinical experience. The X-rays-text paired dataset enables multimodal DL models to capture richer data features from multiple dimensions, even though the PFNF X-rays are insufficient. Multimodal DL model exhibits superior performance to unimodal DL model.

III