Hyug-Gi Kim

Background/Objectives: Accurate visual categorization of small enhancing lesions on contrast-enhanced breast MRI remains challenging. The purpose of this study was to develop a deep learning (DL) model for detecting small (≤2 cm) invasive breast cancers using multi-institutional breast MRI data. Methods: This retrospective study included a total of 1721 women (mean age: 54 years; range, 21–87 years) with T1-stage invasive breast cancer across five hospitals. All patients underwent dynamic contrast-enhanced breast MRI before surgery between 2010 and 2020. Enhancing breast masses on early postcontrast T1-weighted images (14,917 labeled cancer images and 1443 labeled benign images) were used as the ground truth. In this multicenter cancer-enriched cohort, a DL model was developed to distinguish cancerous lesions from noncancerous lesions. MR images were independently reviewed by three breast imaging radiologists to detect subcentimeter (≤1 cm) cancers, and the readers were informed that each examination contained a single cancer. The performance of the model was evaluated using the area under the precision–recall curve (AUPRC), sensitivity, precision, and F1 score. Results: The DL model achieved an AUPRC of 0.42 in detecting small (≤2 cm) invasive breast cancers presenting as enhancing masses on MRI (sensitivity, 83.2%; precision, 33.2%; F1 score, 0.47). For subcentimeter (≤1 cm) cancers, the detection performance of the model was lower (sensitivity, 75.3%; precision, 20.8%; F1 score, 0.33) than that of the radiologists (mean sensitivity, 89.5%; mean precision, 72.9%; mean F1 score, 0.80). When the radiologists used the DL model, their mean precision in detecting subcentimeter invasive breast cancers improved from 72.9% to 83.2%, with no significant change in sensitivity (89.5% vs. 86.8%, p > 0.05). Conclusions: The DL model demonstrated potential as an assistive tool by improving radiologists’ precision for detecting small invasive breast cancers on contrast-enhanced breast MRI in this enriched reader-study setting, although it did not significantly improve their sensitivity.

The rapid development of deep learning in medical imaging has significantly enhanced the capabilities of artificial intelligence while simultaneously introducing challenges, including the need for vast amounts of training data and the labor-intensive tasks of labeling and segmentation. Generative adversarial networks (GANs) have emerged as a solution, offering synthetic image generation for data augmentation and streamlining medical image processing tasks through models such as cGAN, CycleGAN, and StyleGAN. These innovations not only improve the efficiency of image augmentation, reconstruction, and segmentation, but also pave the way for unsupervised anomaly detection, markedly reducing the reliance on labeled datasets. Our investigation into GANs in medical imaging addresses their varied architectures, the considerations for selecting appropriate GAN models, and the nuances of model training and performance evaluation. This paper aims to provide radiologists who are new to GAN technology with a thorough understanding, guiding them through the practical application and evaluation of GANs in brain imaging with two illustrative examples using CycleGAN and pixel2style2pixel (pSp)-combined StyleGAN. It offers a comprehensive exploration of the transformative potential of GANs in medical imaging research. Ultimately, this paper strives to equip radiologists with the knowledge to effectively utilize GANs, encouraging further research and application within the field.

The purpose of this study was to develop a mammography-based deep learning (DL) model for predicting the risk of breast cancer in Asian women. This retrospective study included 287 examinations in 153 women in the cancer group and 736 examinations in 447 women in the negative group, obtained from the databases of two tertiary hospitals between November 2012 and March 2022. All examinations were labeled as either dense breast or nondense breast, and then randomly assigned to either training, validation, or test sets. DL models, referred to as image-level and examination-level models, were developed. Both models were trained to predict whether or not the breast would develop breast cancer with two datasets: the whole dataset and the dense-only dataset. The performance of DL models was evaluated using the accuracy, precision, sensitivity, specificity, F1 score, and area under the receiver operating characteristic curve (AUC). On a test set, performance metrics for the four scenarios were obtained: image-level model with whole dataset, image-level model with dense-only dataset, examination-level model with whole dataset, and examination-level model with dense-only dataset with AUCs of 0.71, 0.75, 0.66, and 0.67, respectively. Our DL models using mammograms have the potential to predict breast cancer risk in Asian women.

Purpose: Intracranial vertebral artery dissection (VAD) is being increasingly recognized as a leading cause of Wallenberg syndrome and subarachnoid hemorrhage. Conventional angiography is considered the standard diagnostic modality, but the diagnosis of VAD remains challenging. This study aimed to compare the diagnostic performance of high-resolution vessel wall imaging (HR-VWI) with digital subtraction angiography (DSA) for intracranial VAD. Materials and methods: Twenty-four patients with 27 VADs, who underwent both HR-VWI and DSA within 2 weeks, were consecutively enrolled in the study from March 2016 to September 2020. HR-VWI and DSA were performed to diagnose VAD and to categorize its angiographic features as either definite dissection or suspicious dissection. Features of HR-VWI were used to evaluate direct arterial wall imaging. The reference standard was set from the clinicoradiologic diagnosis. Two independent raters evaluated the angiographic features, dissection signs, and interrater agreement. Each subject was also dichotomized into two groups (suspicious or definite VAD) in each modality, and diagnosis from HR-VWI and DSA was compared with the final diagnosis by consensus. Results: HR-VWI had higher agreement (90.6% vs. 53.1%) with the final diagnosis and better interrater reliability (kappa value (κ) = 0.91; 95% confidence interval (CI) = 0.64–1.00) compared with DSA (κ = 0.58; 95% CI = 0.35–1.00). HR-VWI provided a more detailed identification of dissection signs (77.7% vs. 22.2%) and better reliability (κ = 0.88; 95% CI = 0.58–1.00 vs. κ = 0.75; 95% CI = 0.36–1.00), compared to DSA. HR-VWI was comparable to DSA for the depiction of angiographic features for VAD. Conclusions: HR-VWI may be useful to evaluate VAD, with better diagnostic confidence compared to DSA.

Langerhans cell histiocytosis (LCH) is a rare neoplastic disorder characterized by the clonal proliferation of CD1a +/CD 207 + dendritic cells, whose features are similar to those of epidermal Langerhans cells. LCH is more common in children than in adults. Localized osteolytic lesions in the craniofacial bones are the most common manifestations of LCH. However, LCH can also present as a multifocal and multisystem disease with poor prognosis. Locally aggressive LCH needs to be differentiated from various diseases such as osteomyelitis, malignant bone tumors, and soft tissue sarcomas. However, it is difficult to diagnose, since the imaging findings are nonspecific. We report a case of a highly aggressive LCH in the maxilla accompanied by a fluid-fluid level.

Purpose: The hyperintense acute reperfusion marker (HARM) is characterized by the delayed enhancement of the subarachnoid or subpial space observed on postcontrast fluid-attenuated inversion recovery (FLAIR) images, and is considered a cerebral reperfusion marker for various brain disorders, including infarction. In this study, we evaluated the cerebral distribution patterns of HARM for discriminating between an enhancing subacute infarction and an enhancing mass located in the cortex and subcortical white matter. Materials and methods: We analyzed consecutive patients who experienced a subacute ischemic stroke, were hospitalized, and underwent conventional brain magnetic resonance imaging including postcontrast FLAIR within 14 days from symptom onset, as well as those who had lesions corresponding to a clinical sign detected by diffusion-weighted imaging and postcontrast T1-weighted imaging between May 2019 and May 2021. A total of 199 patients were included in the study. Of them, 94 were finally included in the subacute infarction group. During the same period, 76 enhancing masses located in the cortex or subcortical white matter, which were subcategorized as metastasis, malignant glioma, and lymphoma, were analyzed. We analyzed the overall incidence of HARM in subacute ischemic stroke cases, and compared the enhancement patterns between cortical infarctions and cortical masses. Results: Among 94 patients with subacute stroke, 78 patients (83%) presented HARM, and among 76 patients with subcortical masses, 48 patients (63%) presented peripheral rim enhancement. Of 170 subcortical enhancing lesions, 88 (51.8%) showed HARM, and 78 (88.6%) were determined to be subacute infarction. Among 94 patients with subacute stroke, 48 patients (51%) had diffusion restrictions, and HARM was found in 39 patients (81.2%). Of the 46 patients (49%) without diffusion restriction, 39 patients (84.8%) showed HARM. Conclusions: The presence of HARM was significantly associated with subacute infarctions. For the masses, a peripheral rim enhancement pattern was observed around the mass rather than the cerebral sulci on postcontrast FLAIR.