Machine Learning for Adrenal Gland Segmentation and Classification of Normal and Adrenal Masses at CT

Machine Learning Model with Computed Tomography Radiomics and Clinicobiochemical Characteristics Predict the Subtypes of Patients with Primary Aldosteronism

RATIONALE AND OBJECTIVES

Adrenal venous sampling (AVS) is the primary method for differentiating between primary aldosterone (PA) subtypes. The aim of study is to develop prediction models for subtyping of patients with PA using computed tomography (CT) radiomics and clinicobiochemical characteristics associated with PA.

MATERIALS AND METHODS

This study retrospectively enrolled 158 patients with PA who underwent AVS between January 2014 and March 2021. Neural network machine learning models were developed using a two-stage analysis of triple-phase abdominal CT and clinicobiochemical characteristics. In the first stage, the models were constructed to classify unilateral or bilateral PA; in the second stage, they were designed to determine the predominant side in patients with unilateral PA. The final proposed model combined the best-performing models from both stages. The model's performance was evaluated using repeated stratified five-fold cross-validation. We employed paired t-tests to compare its performance with the conventional imaging evaluations made by radiologists, which categorize patients as either having bilateral PA or unilateral PA on one side.

RESULTS

In the first stage, the integrated model that combines CT radiomic and clinicobiochemical characteristics exhibited the highest performance, surpassing both the radiomic-alone and clinicobiochemical-alone models. It achieved an accuracy and F1 score of 80.6% ± 3.0% and 74.8% ± 5.2% (area under the receiver operating curve [AUC] = 0.778 ± 0.050). In the second stage, the accuracy and F1 score of the radiomic-based model were 88% ± 4.9% and 81.9% ± 6.2% (AUC=0.831 ± 0.087). The proposed model achieved an accuracy and F1 score of 77.5% ± 3.9% and 70.5% ± 7.1% (AUC=0.771 ± 0.046) in subtype diagnosis and lateralization, surpassing the accuracy and F1 score achieved by radiologists' evaluation (p < .05).

CONCLUSION

The proposed machine learning model can predict the subtypes and lateralization of PA. It yields superior results compared to conventional imaging evaluation and has potential to supplement the diagnostic process in PA.

Development and validation of machine-learning models for the difficulty of retroperitoneal laparoscopic adrenalectomy based on radiomics

Objective

The aim is to construct machine learning (ML) prediction models for the difficulty of retroperitoneal laparoscopic adrenalectomy (RPLA) based on clinical and radiomic characteristics and to validate the models.

Methods

Patients who had undergone RPLA at Shanxi Bethune Hospital between August 2014 and December 2020 were retrospectively gathered. They were then randomly split into a training set and a validation set, maintaining a ratio of 7:3. The model was constructed using the training set and validated using the validation set. Furthermore, a total of 117 patients were gathered between January and December 2021 to form a prospective set for validation. Radiomic features were extracted by drawing the region of interest using the 3D slicer image computing platform and Python. Key features were selected through LASSO, and the radiomics score (Rad-score) was calculated. Various ML models were constructed by combining Rad-score with clinical characteristics. The optimal models were selected based on precision, recall, the area under the curve, F1 score, calibration curve, receiver operating characteristic curve, and decision curve analysis in the training, validation, and prospective sets. Shapley Additive exPlanations (SHAP) was used to demonstrate the impact of each variable in the respective models.

Results

After comparing the performance of 7 ML models in the training, validation, and prospective sets, it was found that the RF model had a more stable predictive performance, while xGBoost can significantly benefit patients. According to SHAP, the variable importance of the two models is similar, and both can reflect that the Rad-score has the most significant impact. At the same time, clinical characteristics such as hemoglobin, age, body mass index, gender, and diabetes mellitus also influenced the difficulty.

Conclusion

This study constructed ML models for predicting the difficulty of RPLA by combining clinical and radiomic characteristics. The models can help surgeons evaluate surgical difficulty, reduce risks, and improve patient benefits.

Multi-organ segmentation of CT via convolutional neural network: impact of training setting and scanner manufacturer

Objective. Automated organ segmentation on CT images can enable the clinical use of advanced quantitative software devices, but model performance sensitivities must be understood before widespread adoption can occur. The goal of this study was to investigate performance differences between Convolutional Neural Networks (CNNs) trained to segment one (single-class) versus multiple (multi-class) organs, and between CNNs trained on scans from a single manufacturer versus multiple manufacturers.Methods. The multi-class CNN was trained on CT images obtained from 455 whole-body PET/CT scans (413 for training, 42 for testing) taken with Siemens, GE, and Phillips PET/CT scanners where 16 organs were segmented. The multi-class CNN was compared to 16 smaller single-class CNNs trained using the same data, but with segmentations of only one organ per model. In addition, CNNs trained on Siemens-only (N = 186) and GE-only (N = 219) scans (manufacturer-specific) were compared with CNNs trained on data from both Siemens and GE scanners (manufacturer-mixed). Segmentation performance was quantified using five performance metrics, including the Dice Similarity Coefficient (DSC).Results. The multi-class CNN performed well compared to previous studies, even in organs usually considered difficult auto-segmentation targets (e.g., pancreas, bowel). Segmentations from the multi-class CNN were significantly superior to those from smaller single-class CNNs in most organs, and the 16 single-class models took, on average, six times longer to segment all 16 organs compared to the single multi-class model. The manufacturer-mixed approach achieved minimally higher performance over the manufacturer-specific approach.Significance. A CNN trained on contours of multiple organs and CT data from multiple manufacturers yielded high-quality segmentations. Such a model is an essential enabler of image processing in a software device that quantifies and analyzes such data to determine a patient's treatment response. To date, this activity of whole organ segmentation has not been adopted due to the intense manual workload and time required.

Artificial Intelligence in Minimally Invasive Adrenalectomy: Using Deep Learning to Identify the Left Adrenal Vein

BACKGROUND

Minimally invasive adrenalectomy is the main surgical treatment option for the resection of adrenal masses. Recognition and ligation of adrenal veins are critical parts of adrenal surgery. The utilization of artificial intelligence and deep learning algorithms to identify anatomic structures during laparoscopic and robot-assisted surgery can be used to provide real-time guidance.

METHODS

In this experimental feasibility study, intraoperative videos of patients who underwent minimally invasive transabdominal left adrenalectomy procedures between 2011 and 2022 in a tertiary endocrine referral center were retrospectively analyzed and used to develop an artificial intelligence model. Semantic segmentation of the left adrenal vein with deep learning was performed. To train a model, 50 random images per patient were captured during the identification and dissection of the left adrenal vein. A randomly selected 70% of data was used to train models while 15% for testing and 15% for validation with 3 efficient stage-wise feature pyramid networks (ESFPNet). Dice similarity coefficient (DSC) and intersection over union scores were used to evaluate segmentation accuracy.

RESULTS

A total of 40 videos were analyzed. Annotation of the left adrenal vein was performed in 2000 images. The segmentation network training on 1400 images was used to identify the left adrenal vein in 300 test images. The mean DSC and sensitivity for the highest scoring efficient stage-wise feature pyramid network B-2 network were 0.77 (±0.16 SD) and 0.82 (±0.15 SD), respectively, while the maximum DSC was 0.93, suggesting a successful prediction of anatomy.

CONCLUSIONS

Deep learning algorithms can predict the left adrenal vein anatomy with high performance and can potentially be utilized to identify critical anatomy during adrenal surgery and provide real-time guidance in the near future.

Age-related morphometrics of normal adrenal glands based on deep learning-aided segmentation

OBJECTIVE

This study aims to evaluate the morphometrics of normal adrenal glands in adult patients semiautomatically using a deep learning-based segmentation model.

MATERIALS AND METHODS

A total of 520 abdominal CT image series with normal findings, from January 1, 2016, to March 14, 2019, were retrospectively collected for the training of the adrenal segmentation model. Then, 1043 portal venous phase image series of inpatient contrast-enhanced abdominal CT examinations with normal adrenal glands were included for analysis and grouped by every 10-year gap. A 3D U-Net-based segmentation model was used to predict bilateral adrenal labels followed by manual modification of labels as appropriate. Quantitative parameters (volume, CT value, and diameters) of the bilateral adrenal glands were then analyzed.

RESULTS

In the study cohort aged 18-77 years old (554 males and 489 females), the left adrenal gland was significantly larger than the right adrenal gland [all patients, 2867.79 (2317.11-3499.89) mm³ vs. 2452.84 (1983.50-2935.18) mm³, P < 0.001]. Male patients showed a greater volume of bilateral adrenal glands than females in all age groups (all patients, left: 3237.83 ± 930.21 mm³ vs. 2646.49 ± 766.42 mm³, P < 0.001; right: 2731.69 ± 789.19 mm³ vs. 2266.18 ± 632.97 mm³, P = 0.001). Bilateral adrenal volume in male patients showed an increasing then decreasing trend as age increased that peaked at 38-47 years old (left: 3416.01 ± 886.21 mm³, right: 2855.04 ± 774.57 mm³).

CONCLUSIONS

The semiautomated measurement revealed that the adrenal volume differs as age increases. Male patients aged 38-47 years old have a peaked adrenal volume.

Automated Adrenal Gland Disease Classes Using Patch-Based Center Symmetric Local Binary Pattern Technique with CT Images

Incidental adrenal masses are seen in 5% of abdominal computed tomography (CT) examinations. Accurate discrimination of the possible differential diagnoses has important therapeutic and prognostic significance. A new handcrafted machine learning method has been developed for the automated and accurate classification of adrenal gland CT images. A new dataset comprising 759 adrenal gland CT image slices from 96 subjects were analyzed. Experts had labeled the collected images into four classes: normal, pheochromocytoma, lipid-poor adenoma, and metastasis. The images were preprocessed, resized, and the image features were extracted using the center symmetric local binary pattern (CS-LBP) method. CT images were next divided into 16 × 16 fixed-size patches, and further feature extraction using CS-LBP was performed on these patches. Next, extracted features were selected using neighborhood component analysis (NCA) to obtain the most meaningful ones for downstream classification. Finally, the selected features were classified using k-nearest neighbor (kNN), support vector machine (SVM), and neural network (NN) classifiers to obtain the optimum performing model. Our proposed method obtained an accuracy of 99.87%, 99.21%, and 98.81% with kNN, SVM, and NN classifiers, respectively. Hence, the kNN classifier yielded the highest classification results with no pathological image misclassified as normal. Our developed fixed patch CS-LBP-based automatic classification of adrenal gland pathologies on CT images is highly accurate and has low time complexity [Formula: see text]. It has the potential to be used for screening of adrenal gland disease classes with CT images.