3. Results
3.1 Clinical Characteristics of the Study Population
After screening, we enrolled 878 eligible patients in the training (n =
702) and validation (n = 176) cohorts of the classification model. There
were no significant between-cohort differences in the general
information.
3.2 Results of the Semantic Segmentation Models
After training for 20 epochs, the performance of U-net and Deeplabv3
quickly stabilized; moreover, they had Dice coefficients of 0.953 and
0.961, respectively (Figure 4). Deeplabv3 performed slightly better than
U-net. We used the Deeplabv3 network trained to segment the sinus region
in CT images.
3.3 Performance of the Classification Models
When trained using a single image as a unit, the areas under the curve
(AUCs) of the efficientnet_b0, resnet50, inception_resnet_v2, and
Xception networks were 0.84, 0.86, 0.83, and 0.86, respectively. When
trained using each patient as a unit, the AUCs of the four neural
networks increased to 0.89, 0.90, 0.88, and 0.90, respectively (Figure
5). The accuracy of the training process reflects the overall
classification performance of the DL models (Figure: 6).
We incorporated a confusion matrix for class-wise comparison to evaluate
whether the model could reliably detect and classify objects. As shown
in Figures 7 and 8, the specificity and sensitivity of the four networks
were higher when using each patient, rather than single images, as
units.
The presentation of Grad-Cams allows elucidation of how the DL network
captured image features for prediction and resolves doubts regarding the
network’s ability to learn in the appropriate direction. Yellow areas
shown in the Grad-Cams had the strongest correlation with the
classification. For patients with ECRS, the yellow areas represent
characteristics associated with an increased risk of ECRS. The
efficientnet_b0 network was used as an example. Figures 9, 10, 11, and
12 show Grad-Cams for both the bone window and soft tissue window images
of the patients. For both diseases, yellow areas were concentrated in
areas with lesions, which is consistent with our medical experience.