To construct an intelligent segmentation and T-stage diagnostic model for esophageal cancer based on the DAEUnet and ConvNeXt networks using transfer learning.

Dicom raw data from 126 patients diagnosed with esophageal cancer between January 2018 and April 2022 were collected, including 100 cases from Department of Thoracic Surgery at the First Affiliated Hospital of Army Medical University and 26 cases from the Department of Thoracic Surgery at Shanxi Cancer Hospital. After data augmentation, a total of 60275 images were obtained. The DAEUnet esophageal cancer intelligent segmentation network was built, and on this basis, 3 classification networks, ConvNeXt, Swin Transformer, and ResNet were constructed for T-stage diagnosis of esophageal cancer.

The Dice similarity coefficient (DSC) for esophageal cancer intelligent segmentation using the DAEUnet network was 0.82, and the DSC value of the esophagus, aorta, normal esophagus, mediastinal lymph nodes, and heart was 72.4%, 87.5%, 79.3%, 60.5% and 96.8%, respectively. Among the 3 T-stage diagnosis models for esophageal cancer, the ConvNeXt model performed the best, with a precision value for T1~T4 stages of 0.65, 0.727, 0.889 and 0.92, respectively, and an AUC value of 0.892, which were superior to the ResNet and Swin Transformer networks.

The proposed DAEUnet and ConvNeXt-based intelligent segmentation and T-stage diagnosis model for esophageal cancer improves T-stage accuracy and treatment efficiency.

To establish and validate a multimodal MRI radiomics model based on intratumoral and peritumoral heterogeneity for prediction of spatial pattern of locally recurrent high-grade gliomas (HGGs).

A retrospective analysis was conducted on the clinical and imaging data of all HGGs patients who underwent maximum safe resection followed by postoperative radiotherapy combined with temozolomide treatment and experienced in local recurrence in Army Medical Center of PLA from 2012 to 2021. Two radiologists independently assessed the spatial patterns of locally recurrence HGGs through continuous follow-up MRI data, and primarily categorized the pattern into intra-resection cavity recurrence and extra-resection cavity recurrence. The subjected patients were randomly divided into a training set and a validation set in a 7∶3 ratio. In the training set, Pearson or Spearman correlation analysis and least absolute shrinkage and selection operator (LASSO) analysis were employed to screen radiomic features within the intratumoral and peritumoral regions, as well as to calculate radiomic scores. A radiomics model was established using logistic regression analysis. The performance of the model was assessed using calibration curves, Hosmer-Lemeshow goodness-of-fit test, and the area under the receiver operating characteristic curve (AUC). Validation of the model was performed in the validation set.

A total of 121 patients with locally recurrent HGGs were enrolled in this study, including 54 in intra-resection cavity recurrence group and 67 in extra-resection cavity recurrence group. Among them, 84 were assigned into the training set and 37 into the validation set. In the training set, the radiomics score for the extra-resection cavity recurrence group was 0.424 (0.278, 0.573), which was higher than that for the intra-resection cavity recurrence group [-0.030 (-0.226,0.248), P<0.001]. In the validation set, the radiomics score for the extra-resection cavity recurrence group was 0.369 (0.258, 0.487), which was higher than that for the intra-resection cavity recurrence group [0.277 (0.103, 0.322), P=0.033]. The established radiomics model exhibited good calibration and performed well in predicting spatial recurrence patterns, with an AUC value of 0.844 (95%CI: 0 749~0.914) in the training set and 0.706 (95%CI: 0.534~0.844) in the validation set.

Our multimodal radiomics model combined with intratumoral and peritumoral heterogeneity can predict the spatial pattern of locally recurrent HGGs, providing a basis for individualized treatment of HGGs.

To evaluate the correction rate, accuracy of pedicle screw fixation and overall clinical efficacy of intravertebral osteotomy and internal fixation surgery with the assistance of 3D printing guide plates in treatment of severe kyphosis.

A single-center nonrandomized clinical pilot study was conducted on 19 patients (8 males and 11 females) with severe kyphosis undergoing intravertebral osteotomy between December 2018 and June 2023. Seven of them (CAD group) had preoperative planning with computer-aided design (CAD) and intraoperative guidance of individualized 3D printing guide plates. And another 12 patients (control group) were corrected with conventional pedicle screw placement. Postoperative evaluation included assessment of posterior Cobb angle, spinal angular correction rate, accuracy of pedicle screw placement and Oswestry Dysfunction Index (ODI) questionnaire.

The 19 patients were at a mean age of 48.0 years, and followed up for 26.4 (9~54) months. All of them achieved relatively satisfactory corrective results, with those of the CAD group having a correction rate of 96.83% and those of the control group of 86.61%. There were no statistical differences in average intraoperative blood loss (857 vs 1 045 mL) and average operative time (344 vs 402 min), but significant difference was observed in average length of hospital stay (11 vs 18 d, P<0.05) between the 2 groups. A total of 278 nails were placed in this study, including 70 guide-assisted pedicle screws, 97.1% of which were grade A or B. In the control group, 208 pedicle screws were placed, 93.8% of which were grade A or B. Postoperative CT/X-ray scanning displayed that both groups achieved certain correction for kyphosis. No obvious difference was found in the average spinal angular correction (43.37° vs 36.10°), and significantly higher correction rate was seen in the CAD group than the control group (96.83% vs 86.61%, P<0.01). The ODI value was notably lower in the CAD group than the control group (P<0.05).

CAD-assisted preoperative planning, surgical simulation and individualized 3D printing guide plates can promote surgical correction and accuracy of pedicle screw placement and improves the quality of life of patients with severe kyphotic deformity.

To establish an artificial intelligence (AI) diagnostic model for the histopathologic diagnosis of extramammary Paget's disease (EMPD) and to evaluate its efficiency for the diagnosis and differential diagnosis of EMPD.

All non-tumor skin disease patients who underwent skin tissue biopsy in Department of Dermatology of First Affiliated Hospital of Army Medical University from September 2003 to February 2023 were recruited, and their pathological data were collected, including EMPD, Bowen's disease (BD), squamous cell carcinoma (SCC), and epidermal hyperplasia and hypertrophy. With EMPD as the main research subject, the histopathological images of BD, SCC, and non-tumor skin diseases were included in the study. The histopathological data of 4 types of diseases was classified and diagnosed by ResNet101 and DenseNet121 deep learning neural networks, and the performance of these models was evaluated.

The AUC values of the ResNet101 diagnostic model for the diagnosis of EMPD, BD, SCC and non-tumor skin diseases on the images at ×20 magnification were 0.97, 0.98, 1.00 and 0.96, respectively, with an accuracy of 0.925±0.011, while the AUC values on the images at ×40 magnification were 1.00, 0.99, 1.00 and 0.97, respectively, with an accuracy of 0.943±0.017. The AUC values of the DenseNet121 diagnostic model for the diagnosis of 4 diseases on the images at ×20 magnification were 0.98, 0.95, 0.99 and 1.00, respectively, with an accuracy of 0.912±0.034, while the AUC values on the images at ×40 magnification were 0.99, 0.96, 1.00 and 1.00, respectively, with an accuracy of 0.971±0.012. Our results indicated that the histopathologic diagnostic model could effectively differentiate EMPD from BD, SCC and non-tumor skin diseases at low power magnification. The FLPOs of ResNet101 was 786.6 M, and the parameter was 4.5 M; The FLPOs of DensNet121 was 289.7 M, and the parameter was 0.8M.

Our AI diagnostic model is of good effectiveness in the diagnosis and differential diagnosis of EMPD. DenseNet121 is recommended as the dermatopathological diagnostic model of this study.

To create an intelligent segmentation model of deep learning based on the magnetic resonance（MR）images of the prostate zones.

The T2WI sequence data of 33 patients with prostate cancer, including 6 at T1 stage, 15 at T2 stage, 9 at T3 stage, and 3 at T4 stage, receiving MRI scanning in Shanxi Provincial Cancer Hospital from January 2018 to October 2020 were collected. While, the T2WI sequence data of 346 patients with prostate cancer from the Prostate MRI Public Dataset provided by Radboud University Nijmegen Medical Centre, Netherlands were also collected. Then these totally 379 cases were randomly divided into training sets, validation sets and test sets at a ratio of 7∶1∶2（265, 38 and 76 cases, respectively）. Based on Unet model, Vgg16 was used as the encoder, and multiple layers of convolution were used with the transfer learning strategy at the same time to build a Vgg16-Unet model. Then with a gold standard of the prostate zones（anterior fibrous matrix zone, central zone, peripheral zone, and transition zone）manually segmented by doctors, Dice similarity coefficient（DSC）and 95% Hausdorff surface distance（HD95）were employed to evaluate the segmentation accuracy of the prostate zones on the test set.

The model achieved relatively higher accuracy of segmentation for anterior fibrous matrix zone, central zone, peripheral zone and transition zone on the test set. The average DSC was 56.95%, 47.28%, 80.78%, and 90.63%, and the average HD95 value was 20.84, 20.02, 15.39 and 11.20 mm, respectively. The volume predicted by the model was consistent with the volume measured by manual segmentation, and their differences were basically within the 95% consistency interval.

The segmentation accuracy of our Vgg16-Unet model is better than that of 3 classical variants, Unet, Unet++ and ResUnet++. The model can significantly improve the segmentation efficiency of prostate cancer MRI images and reduce the workload of clinicians.

To explore the advantages of Swin-Transformer（SwinT）network in the intelligent diagnosis of prostate cancer（PCA）based on MRI images by comparing Dense-Net, Res-Net, and Vision-Transformer（ViT）networks.

A total of 3 017 MRI images including T2WI and fT2WI from 152 patients with PCA confirmed by puncture biopsy were retrospectively collected from Shanxi Cancer Hospital and Second Affiliated Hospital of Army Medical University between April 2020 and March 2022. According to the results of clinical T stage, these images were divided into low- to middle-risk group（T≤T2c）and high-risk group（T≥T3a）. And then, the patients were randomly divided into training set（n=107）, verification set（n=15）and test set（n=30）at a ratio of 7:1:2 to train the intelligent T-stage diagnosis models. Finally, accuracy, precision, confusion matrix, receiver operating characteristic（ROC）curves and areas under the curve（AUC）were used to evaluate the performance of intelligent T-stage diagnosis of each network.

In the low- to middle-risk and high-risk groups, the accuracy and AUC of above 4 networks were 0.587 and 0.630, 0.410 and 0.477, 0.600 and 0.648, and 0.680 and 0.708, respectively. The Grad-CAM of SwinT networks had the attention almost focused on the prostate, showing the best feature extraction.

Compared with the models based on Dense-Net, Res-Net and ViT networks, the SwinT model achieves the best predictive performance in the task of classification of PCA MRI images. The model can be used for the automatic diagnosis of PCA T-stage, and is helpful to improve diagnostic efficiency.