Development and Evaluation of Deep Learning-based Automated Segmentation of Pituitary Adenoma in Clinical Task

Impact of growth hormone-secreting pituitary adenoma on limbic system and its correlation with cognitive impairment

Purpose

To assess the quantitative gray matter volume of the limbic system in growth hormone-secreting pituitary adenoma (GHPAs) patients and its correlation to cognitive function.

Method

91 right-handed patients with pituitary adenomas were retrospectively included from the First Affiliated Hospital of Sun Yat-sen University -48 with GHPAs and 43 with non-functioning pituitary adenomas (NFPAs). Participants underwent serum hormone assessment, regular sellar MRI scanning with T1WI-MPRAGE. Cognitive function was gauged using MoCA and MMSE. Brain region auto-segmentation and gray matter volume calculation were conducted on the Brainsite platform.

Results

Compared to NFPAs patients, GHPAs patients had higher gray matter volume (758,285 vs 674,610 mm³, p < 0.001). No significant volumetric differences in both sides of limbic system gray matter while there were evident differences in the relative volumes of limbic system gray matter between groups. GHPAs patients scored lower on MOCA (24.0 (2.18) vs 25.1 (2.28), p < 0.031), with no difference in MMSE. We observed a significant correlation between the relative limbic volume and MOCA scales, while no evident correlation was found between relative limbic volume and serum hormone or tumor aggressiveness. Univariate and multivariate Logistic regression showed that hippocampus and limbic cortex (parahippocampal gyrus and internal olfactory area) of advantageous hemisphere correlated significantly with occurrence of mild cognitive impairment with the C-statistic reaching 0.90.

Conclusion

Patients with GHPAs show a relative decrease in limbic gray matter volume, especially in the hippocampus and limbic cortex of the dominant hemisphere, which is associated with mild cognitive impairment.

Development and validation of a prediction model for consistency of pituitary adenoma: the PiTCon score

PURPOSE

Pituitary adenomas (PAs) usually have a soft consistency, facilitating gross total resection. However, 5-13% of PAs with fibrous consistency are challenging to remove entirely and are accompanied by greater morbimortality. This study aims to identify the clinical and radiological characteristics that correlate with PA fibrous consistency preoperatively. A simple scoring system has been proposed to predict incidence of fibrous PAs.

MATERIALS AND METHODS

Consecutive interventions (226) were analyzed, all performed through an endoscopic endonasal transsphenoidal approach. Univariable and multivariable logistic regression analysis was performed. Hosmer-Lemeshow test and receiver operating characteristic (ROC) curves were assessed to evaluate the model. A point scoring system (PiTCon) was derived based on the multivariable regression model. Our study aimed to identify the clinical and radiological characteristics that correlate with fibrous tumor consistency preoperatively.

RESULTS

The best diagnostic accuracy for predicting PA consistency consisted of five predictive factors: age, compressive symptoms, panhypopituitarism, craniocaudal extension of the PA in mm, and prior surgery. The multivariable model achieved good discrimination with an area under the curve (AUC) of the ROC curve being 0.82 and the 95% CI 0.76 to 0.88. Internal validation yielded an optimism-adjusted C-statistic of 0.80 (95% CI 0.74 to 0.86). A point scoring system (PiTCon score) was designed using the best predictive model.

CONCLUSIONS

PA consistency can be estimated preoperatively regarding clinical and radiological characteristics. We propose a point-based scoring system (PiTCon score) that can better guide neurosurgeons in clinical decision-making and surgical risk assessment and help establish and describe patient prognosis.

deepPGSegNet: MRI-based pituitary gland segmentation using deep learning

Introduction

In clinical research on pituitary disorders, pituitary gland (PG) segmentation plays a pivotal role, which impacts the diagnosis and treatment of conditions such as endocrine dysfunctions and visual impairments. Manual segmentation, which is the traditional method, is tedious and susceptible to inter-observer differences. Thus, this study introduces an automated solution, utilizing deep learning, for PG segmentation from magnetic resonance imaging (MRI).

Methods

A total of 153 university students were enrolled, and their MRI images were used to build a training dataset and ground truth data through manual segmentation of the PGs. A model was trained employing data augmentation and a three-dimensional U-Net architecture with a five-fold cross-validation. A predefined field of view was applied to highlight the PG region to optimize memory usage. The model's performance was tested on an independent dataset. The model's performance was tested on an independent dataset for evaluating accuracy, precision, recall, and an F1 score.

Results and discussion

The model achieved a training accuracy, precision, recall, and an F1 score of 92.7%, 0.87, 0.91, and 0.89, respectively. Moreover, the study explored the relationship between PG morphology and age using the model. The results indicated a significant association between PG volume and midsagittal area with age. These findings suggest that a precise volumetric PG analysis through an automated segmentation can greatly enhance diagnostic accuracy and surveillance of pituitary disorders.

Preoperatively Predicting Ki67 Expression in Pituitary Adenomas Using Deep Segmentation Network and Radiomics Analysis Based on Multiparameter MRI

RATIONALE AND OBJECTIVES

Ki67 proliferation index is associated with more aggressive tumor behavior and recurrence of pituitary adenomas (PAs). Recently, radiomics and deep learning have been introduced into the study of pituitary tumors. The present study aimed to investigate the feasibility of predicting the Ki67 proliferation index of PAs using the deep segmentation network and radiomics analysis based on multiparameter MRI.

MATERIALS AND METHODS

First, the cfVB-Net autosegmentation model was trained; subsequently, its performance was evaluated in terms of the dice similarity coefficient (DSC). In the present study, 1214 patients were classified into the high Ki67 expression group (HG) and the low Ki67 expression group (LG). Analyses of three classification models based on radiomics features were performed to distinguish HG from LG. Clinical factors, imaging features, and Radscores were collectively used to create a nomogram in order to effectively predict Ki67 expression.

RESULTS

The cfVB-Net segmentation model demonstrated good performance (DSC: 0.723-0.930). Overall, 18, 15, and 11 optimal features in contrast-enhanced (CE) T1WI, T1WI, and T2WI were obtained for differentiating between HG and LG, respectively. Notably, the best results were presented in the bagging decision tree when CE T1WI and T1WI were combined (area under the receiver operating characteristic curve: training set, 0.927; validation set, 0.831; and independent testing set, 0.825). In the nomogram, age, Hardy' grade, and Radscores were identified as risk predictors of high Ki67 expression.

CONCLUSION

The deep segmentation network and radiomics analysis based on multiparameter MRI exhibited good performance and clinical application value in predicting the expression of Ki67 in PAs.

Predicting delayed remission in Cushing's disease using radiomics models: a multi-center study

Purpose

No multi-center radiomics models have been built to predict delayed remission (DR) after transsphenoidal surgery (TSS) in Cushing's disease (CD). The present study aims to build clinical and radiomics models based on data from three centers to predict DR after TSS in CD.

Methods

A total of 122 CD patients from Peking Union Medical College Hospital, Xuanwu Hospital, and Fuzhou General Hospital were enrolled between January 2000 and January 2019. The T1-weighted gadolinium-enhanced MRI images and clinical data were used as inputs to build clinical and radiomics models. The regions of interest (ROI) of MRI images were automatically defined by a deep learning algorithm developed by our team. The area under the curve (AUC) of receiver operating characteristic (ROC) curves was used to evaluate the performance of the models. In total, 10 machine learning algorithms were used to construct models.

Results

The overall DR rate is 44.3% (54/122). According to multivariate Logistic regression analysis, patients with higher BMI and lower postoperative cortisol levels are more likely to achieve a higher rate of delayed remission. Among the 10 models, XGBoost achieved the best performance among all models in both clinical and radiomics models with AUC values of 0.767 and 0.819 respectively. The results from SHAP value and LIME algorithms revealed that postoperative cortisol level (PoC) and BMI were the most important features associated with DR.

Conclusion

Radiomics models can be built as an effective noninvasive method to predict DR and might be useful in assisting neurosurgeons in making therapeutic plans after TSS for CD patients. These results are preliminary and further validation in a larger patient sample is needed.

Anterior-posterior diameter is a key driver of resectability and complications for pituitary adenomas with suprasellar extension in endoscopic transsphenoidal surgery

BACKGROUND

As endoscopic transsphenoidal approaches are more routinely selected for progressively larger pituitary adenomas with parasellar extension, understanding potential anatomical factors that limit resection and contribute to complications is becoming increasingly important for tailoring a surgical approach. This study aimed to reevaluate existing predictive tools for resectability in pituitary adenomas specifically with suprasellar extension, and furthermore identify any additional measurable features that may be more useful in preoperative planning.

METHODS

A single-center retrospective chart review of adult patients who underwent endoscopic transsphenoidal surgery for pituitary adenomas with suprasellar extension from 2015 to 2020 was performed. Preoperative MRIs were systematically assessed to assign a Knosp classification, a Zurich Pituitary Score (ZPS), and for dimensional measurements of the suprasellar aspect of the lesions. Univariate comparisons and multivariate regression models were employed to assess the influence of these factors on extent of resection and postoperative complications.

RESULTS

Of the 96 patients with suprasellar pituitary adenomas who underwent endoscopic transsphenoidal surgery, 74 patients (77%) had a gross total resection (GTR). Neither Knosp grade nor ZPS score, even when dichotomized, demonstrated an association with GTR (Knosp 3A-4 versus Knosp 0-2, p = 0.069; ZPS III-IV versus ZPS I-II, p = 0.079). Multivariate regression analysis identified suprasellar anterior-posterior tumor diameter (SSAP) as the only significant predictor of extent of resection in this cohort (OR 0.951, 95% CI 0.905-1.000, p = 0.048*). A higher SSAP also had the strongest association with intraoperative CSF leaks (p = 0.0012*) and an increased overall rate of postoperative complications (p = 0.002*). Further analysis of the regression model for GTR suggested an optimal cut point value for SSAP of 23.7 mm, above which predictability for failing to achieve GTR carried a sensitivity of 89% and a specificity of 41%.

CONCLUSIONS

This study is unique in its examination of endoscopic transsphenoidal surgical outcomes for pituitary adenomas with suprasellar extension. Our findings suggest that previously established grading systems based on lateral extension into the cavernous sinus lose their predictive value in lesions with suprasellar extension and, more specifically, with increasing suprasellar anterior-posterior diameter.

Current and Future Advances in Surgical Therapy for Pituitary Adenoma

The vital physiological role of the pituitary gland, alongside its proximity to critical neurovascular structures, means that pituitary adenomas can cause significant morbidity or mortality. While enormous advancements have been made in the surgical care of pituitary adenomas, numerous challenges remain, such as treatment failure and recurrence. To meet these clinical challenges, there has been an enormous expansion of novel medical technologies (eg, endoscopy, advanced imaging, artificial intelligence). These innovations have the potential to benefit each step of the patient's journey, and ultimately, drive improved outcomes. Earlier and more accurate diagnosis addresses this in part. Analysis of novel patient data sets, such as automated facial analysis or natural language processing of medical records holds potential in achieving an earlier diagnosis. After diagnosis, treatment decision-making and planning will benefit from radiomics and multimodal machine learning models. Surgical safety and effectiveness will be transformed by smart simulation methods for trainees. Next-generation imaging techniques and augmented reality will enhance surgical planning and intraoperative navigation. Similarly, surgical abilities will be augmented by the future operative armamentarium, including advanced optical devices, smart instruments, and surgical robotics. Intraoperative support to surgical team members will benefit from a data science approach, utilizing machine learning analysis of operative videos to improve patient safety and orientate team members to a common workflow. Postoperatively, neural networks leveraging multimodal datasets will allow early detection of individuals at risk of complications and assist in the prediction of treatment failure, thus supporting patient-specific discharge and monitoring protocols. While these advancements in pituitary surgery hold promise to enhance the quality of care, clinicians must be the gatekeepers of the translation of such technologies, ensuring systematic assessment of risk and benefit prior to clinical implementation. In doing so, the synergy between these innovations can be leveraged to drive improved outcomes for patients of the future.

Radiomic analysis of preoperative magnetic resonance imaging for the prediction of pituitary adenoma consistency

BACKGROUND

The consistency of pituitary adenomas affects the course of surgical treatment.

PURPOSE

To evaluate the diagnostic capabilities of radiomics based on T1-weighted (T1W) and T2-weighted (T2W) magnetic resonance imaging (MRI) in conjunction with two machine-learning (ML) techniques (support vector machine [SVM] and random forest classifier [RFC]) for assessing the consistency of pituitary adenomas.

MATERIAL AND METHODS

The institutional database was retrospectively scanned for patients who underwent surgical excision of pituitary adenomas. Surgical notes were accepted as a reference for the adenoma consistency. Radiomics analysis was performed on preoperative coronal 3.0T T1W and T2W images. First- and second-order parameters were calculated. Inter-observer reproducibility was assessed with Spearman's Correlation (ρ) and intra-observer reproducibility was evaluated with the intraclass correlation coefficient (ICC). Least absolute shrinkage and selection operator (LASSO) was used for dimensionality reduction. SVM and RFC were used as ML methods.

RESULTS

A total of 52 patients who produced 206 regions of interest (ROIs) were included. Twenty adenomas that produced 88 ROIs had firm consistency. There was both inter-observer and intra-observer reproducibility. Ten parameters that were based on T2W images with high discriminative power and without correlation were chosen by LASSO. The diagnostic performance of SVM and RFC was as follows: sensitivity = 95.580% and 92.950%, specificity = 83.670% and 88.420%, area under the curve = 0.956 and 0.904, respectively.

CONCLUSION

Radiomics analysis based on T2W MRI combined with various ML techniques, such as SVM and RFC, can provide preoperative information regarding pituitary adenoma consistency with high diagnostic accuracy.

Fully automated imaging protocol independent system for pituitary adenoma segmentation: a convolutional neural network-based model on sparsely annotated MRI

This study aims to develop a fully automated imaging protocol independent system for pituitary adenoma segmentation from magnetic resonance imaging (MRI) scans that can work without user interaction and evaluate its accuracy and utility for clinical applications. We trained two independent artificial neural networks on MRI scans of 394 patients. The scans were acquired according to various imaging protocols over the course of 11 years on 1.5T and 3T MRI systems. The segmentation model assigned a class label to each input pixel (pituitary adenoma, internal carotid artery, normal pituitary gland, background). The slice segmentation model classified slices as clinically relevant (structures of interest in slice) or irrelevant (anterior or posterior to sella turcica). We used MRI data of another 99 patients to evaluate the performance of the model during training. We validated the model on a prospective cohort of 28 patients, Dice coefficients of 0.910, 0.719, and 0.240 for tumour, internal carotid artery, and normal gland labels, respectively, were achieved. The slice selection model achieved 82.5% accuracy, 88.7% sensitivity, 76.7% specificity, and an AUC of 0.904. A human expert rated 71.4% of the segmentation results as accurate, 21.4% as slightly inaccurate, and 7.1% as coarsely inaccurate. Our model achieved good results comparable with recent works of other authors on the largest dataset to date and generalized well for various imaging protocols. We discussed future clinical applications, and their considerations. Models and frameworks for clinical use have yet to be developed and evaluated.

Radiomics and machine learning for predicting the consistency of benign tumors of the central nervous system: A systematic review

PURPOSE

Predicting the consistency of benign central nervous system (CNS) tumors prior to surgery helps to improve surgical outcomes. This review summarizes and analyzes the literature on using radiomics and/or machine learning (ML) for consistency prediction.

METHOD

The Medical Literature Analysis and Retrieval System Online (MEDLINE) database was screened for studies published in English from January 1st 2000. Data was extracted according to the PRISMA guidelines and quality of the studies was assessed in compliance with the Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2).

RESULTS

Eight publications were included focusing on pituitary macroadenomas (n = 5), pituitary adenomas (n = 1), and meningiomas (n = 2) using a retrospective (n = 6), prospective (n = 1), and unknown (n = 1) study design with a total of 763 patients for the consistency prediction. The studies reported an area under the curve (AUC) of 0.71-0.99 for their respective best performing model regarding the consistency prediction. Of all studies, four articles validated their models internally whereas none validated their models externally. Two articles stated making data available on request with the remaining publications lacking information with regard to data availability.

CONCLUSIONS

The research on consistency prediction of CNS tumors is still at an early stage regarding the use of radiomics and different ML techniques. Best-practice procedures regarding radiomics and ML need to be followed more rigorously to facilitate the comparison between publications and, accordingly, the possible implementation into clinical practice in the future.

The Knosp Criteria Revisited: 3-Dimensional Volumetric Analysis as a Predictive Tool for Extent of Resection in Complex Endoscopic Pituitary Surgery

BACKGROUND

The Knosp criteria have been the historical standard for predicting cavernous sinus invasion, and therefore extent of surgical resection, of pituitary macroadenomas. Few studies have sought to reappraise the utility of this tool after recent advances in visualization and modeling of tumors in complex endoscopic surgery.

OBJECTIVE

To evaluate our proposed alternative method, using 3-dimensional (3D) volumetric imaging, and whether it can better predict extent of resection in nonfunctional pituitary adenomas.

METHODS

Patients who underwent endoscopic transsphenoidal resection of pituitary macroadenomas at our institution were reviewed. Information was collected on neurological, endocrine, and visual function. Volumetric segmentation was performed using 3D Slicer software. Relationship of tumor volume, clinical features, and Knosp grade on extent of resection was examined.

RESULTS

One hundred forty patients were identified who had transsphenoidal resection of nonfunctional pituitary adenomas. Macroadenomas had a median volume of 6 cm 3 (IQR 3.4-8.7), and 17% had a unilateral Knosp grade of at least 3B. On multiple logistic regression, only smaller log-transformed preoperative tumor volume was independently associated with increased odds of gross total resection (GTR; odds ratio: 0.27, 95% CI: 0.07-0.89, P < .05) when controlling for tumor proliferative status, age, and sex (area under the curve 0.67). The Knosp criteria did not independently predict GTR in this cohort ( P > .05, area under the curve 0.46).

CONCLUSION

Increasing use of volumetric 3D imaging may better anticipate extent of resection compared with the Knosp grade metric and may have a greater positive predictive value for GTR. More research is needed to validate these findings and implement them using automated methods.

Deep-learning-based automatic segmentation and classification for craniopharyngiomas

Objective

Neuronavigation and classification of craniopharyngiomas can guide surgical approaches and prognostic information. The QST classification has been developed according to the origin of craniopharyngiomas; however, accurate preoperative automatic segmentation and the QST classification remain challenging. This study aimed to establish a method to automatically segment multiple structures in MRIs, detect craniopharyngiomas, and design a deep learning model and a diagnostic scale for automatic QST preoperative classification.

Methods

We trained a deep learning network based on sagittal MRI to automatically segment six tissues, including tumors, pituitary gland, sphenoid sinus, brain, superior saddle cistern, and lateral ventricle. A deep learning model with multiple inputs was designed to perform preoperative QST classification. A scale was constructed by screening the images.

Results

The results were calculated based on the fivefold cross-validation method. A total of 133 patients with craniopharyngioma were included, of whom 29 (21.8%) were diagnosed with type Q, 22 (16.5%) with type S and 82 (61.7%) with type T. The automatic segmentation model achieved a tumor segmentation Dice coefficient of 0.951 and a mean tissue segmentation Dice coefficient of 0.8668 for all classes. The automatic classification model and clinical scale achieved accuracies of 0.9098 and 0.8647, respectively, in predicting the QST classification.

Conclusions

The automatic segmentation model can perform accurate multi-structure segmentation based on MRI, which is conducive to clearing tumor location and initiating intraoperative neuronavigation. The proposed automatic classification model and clinical scale based on automatic segmentation results achieve high accuracy in the QST classification, which is conducive to developing surgical plans and predicting patient prognosis.

Methods of preoperative prediction of pituitary adenoma consistency: a systematic review

This study aims to review the current literature on methods of preoperative prediction of pituitary adenoma consistency. Pituitary adenoma consistency may be a limiting factor for successful surgical removal of tumors. Efforts have been made to investigate the possibility of an accurate assessment of the preoperative consistency to allow for safer and more effective surgery planning. We searched major scientific databases and systematically analyzed the results. A total of 54 relevant articles were identified and selected for inclusion. These studies evaluated methods based on either MRI intensity, enhancement, radiomics, MR elastometry, or CT evaluation. The results of these studies varied widely. Most studies used the average intensity of either T2WI or ADC maps. Firm tumors appeared hyperintense on T2WI, although only 55% of the studies reported statistically significant results. There are mixed reports on ADC values in firm tumors with findings of increased values (28%), decreased values (22%), or no correlation (50%). Multiple contrast enhancement-based methods showed good results in distinguishing between soft and firm tumors. There were mixed reports on the utility of MR elastography. Attempts to develop radiomics and machine learning-based models have achieved high accuracy and AUC values; however, they are prone to overfitting and need further validation. Multiple methods of preoperative consistency assessment have been studied. None demonstrated sufficient accuracy and reliability in clinical use. Further efforts are needed to enable reliable surgical planning.

A Convolutional Neural Network Model for Detecting Sellar Floor Destruction of Pituitary Adenoma on Magnetic Resonance Imaging Scans

Objective

Convolutional neural network (CNN) is designed for image classification and recognition with a multi-layer neural network. This study aimed to accurately assess sellar floor invasion (SFI) of pituitary adenoma (PA) using CNN.

Methods

A total of 1413 coronal and sagittal magnetic resonance images were collected from 695 patients with PAs. The enrolled images were divided into the invasive group (n = 530) and the non-invasive group (n = 883) according to the surgical observation of SFI. Before model training, 100 images were randomly selected for the external testing set. The remaining 1313 cases were randomly divided into the training and validation sets at a ratio of 80:20 for model training. Finally, the testing set was imported to evaluate the model performance.

Results

A CNN model with a 10-layer structure (6-layer convolution and 4-layer fully connected neural network) was constructed. After 1000 epoch of training, the model achieved high accuracy in identifying SFI (97.0 and 94.6% in the training and testing sets, respectively). The testing set presented excellent performance, with a model prediction accuracy of 96%, a sensitivity of 0.964, a specificity of 0.958, and an area under the receptor operator curve (AUC-ROC) value of 0.98. Four images in the testing set were misdiagnosed. Three images were misread with SFI (one with conchal type sphenoid sinus), and one image with a relatively intact sellar floor was not identified with SFI.

Conclusion

This study highlights the potential of the CNN model for the efficient assessment of PA invasion.

Machine Learning for the Detection and Segmentation of Benign Tumors of the Central Nervous System: A Systematic Review

Simple Summary

Machine learning in radiology of the central nervous system has seen many interesting publications in the past few years. Since the focus has largely been on malignant tumors such as brain metastases and high-grade gliomas, we conducted a systematic review on benign tumors to summarize what has been published and where there might be gaps in the research. We found several studies that report good results, but the descriptions of methodologies could be improved to enable better comparisons and assessment of biases.

Abstract

Objectives: To summarize the available literature on using machine learning (ML) for the detection and segmentation of benign tumors of the central nervous system (CNS) and to assess the adherence of published ML/diagnostic accuracy studies to best practice. Methods: The MEDLINE database was searched for the use of ML in patients with any benign tumor of the CNS, and the records were screened according to PRISMA guidelines. Results: Eleven retrospective studies focusing on meningioma (n = 4), vestibular schwannoma (n = 4), pituitary adenoma (n = 2) and spinal schwannoma (n = 1) were included. The majority of studies attempted segmentation. Links to repositories containing code were provided in two manuscripts, and no manuscripts shared imaging data. Only one study used an external test set, which raises the question as to whether some of the good performances that have been reported were caused by overfitting and may not generalize to data from other institutions. Conclusions: Using ML for detecting and segmenting benign brain tumors is still in its infancy. Stronger adherence to ML best practices could facilitate easier comparisons between studies and contribute to the development of models that are more likely to one day be used in clinical practice.

Application of Convolutional Neural Network in the Diagnosis of Cavernous Sinus Invasion in Pituitary Adenoma

Objectives

Convolutional neural network (CNN) is a deep-learning method for image classification and recognition based on a multi-layer NN. In this study, CNN was used to accurately assess cavernous sinus invasion (CSI) in pituitary adenoma (PA).

Methods

A total of 371 patients with PA were enrolled in the retrospective study. The cohort was divided into the invasive (n = 102) and non-invasive groups (n = 269) based on surgically confirmed CSI. Images were selected on the T1-enhanced imaging on MR scans. The cohort underwent a fivefold division of randomized datasets for cross-validation. Then, a tenfold augmented dataset (horizontal flip and rotation) of the training set was enrolled in the pre-trained Resnet50 model for transfer learning. The testing set was imported into the trained model for evaluation. Gradient-weighted class activation mapping (Grad-CAM) was used to obtain the occlusion map. The diagnostic values were compared with different dichotomizations of the Knosp grading system (grades 0-1/2-4, 0-2/3a-4, and 0-3a/3b-4).

Results

Based on Knosp grades, 20 cases of grade 0, 107 cases of grade 1, 82 cases of grade 2, 104 cases of grade 3a, 22 cases of grade 3b, and 36 cases of grade 4 were recorded. The CSI rates were 0%, 3.7%, 18.3%, 37.5%, 54.5%, and 88.9%. The predicted accuracies of the three dichotomies were 60%, 74%, and 81%. The area under the receiver operating characteristic (AUC-ROC) of Knosp grade for CSI prediction was 0.84; the cutoff was 2.5 with a Youden value of 0.62. The accuracies of the CNN model ranged from 0.80 to 0.96, with AUC-ROC values ranging from 0.89 to 0.98. The Grad-CAM saliency maps confirmed that the region of interest of the model was around the sellar region.

Conclusions

We constructed a CNN model with a high proficiency at CSI diagnosis. A more accurate CSI identification was achieved with the constructed CNN than the Knosp grading system.

Development and Validation of a Deep Learning Algorithm to Automatic Detection of Pituitary Microadenoma From MRI

Background: It is often difficult to diagnose pituitary microadenoma (PM) by MRI alone, due to its relatively small size, variable anatomical structure, complex clinical symptoms, and signs among individuals. We develop and validate a deep learning -based system to diagnose PM from MRI.

Methods: A total of 11,935 infertility participants were initially recruited for this project. After applying the exclusion criteria, 1,520 participants (556 PM patients and 964 controls subjects) were included for further stratified into 3 non-overlapping cohorts. The data used for the training set were derived from a retrospective study, and in the validation dataset, prospective temporal and geographical validation set were adopted. A total of 780 participants were used for training, 195 participants for testing, and 545 participants were used to validate the diagnosis performance. The PM-computer-aided diagnosis (PM-CAD) system consists of two parts: pituitary region detection and PM diagnosis. The diagnosis performance of the PM-CAD system was measured using the receiver operating characteristics (ROC) curve and area under the ROC curve (AUC), calibration curve, accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1-score.

Results: Pituitary microadenoma-computer-aided diagnosis system showed 94.36% diagnostic accuracy and 98.13% AUC score in the testing dataset. We confirm the robustness and generalization of our PM-CAD system, the diagnostic accuracy in the internal dataset was 96.50% and in the external dataset was 92.26 and 92.36%, the AUC was 95.5, 94.7, and 93.7%, respectively. In human-computer competition, the diagnosis performance of our PM-CAD system was comparable to radiologists with >10 years of professional expertise (diagnosis accuracy of 94.0% vs. 95.0%, AUC of 95.6% vs. 95.0%). For the misdiagnosis cases from radiologists, our system showed a 100% accurate diagnosis. A browser-based software was designed to assist the PM diagnosis.

Conclusions: This is the first report showing that the PM-CAD system is a viable tool for detecting PM. Our results suggest that the PM-CAD system is applicable to radiology departments, especially in primary health care institutions.