Automating radiology report generation can significantly alleviate radiologists’ workloads. Previous research has primarily focused on realizing highly concise observations while neglecting the precise attributes that determine the severity of diseases (e.g., small pleural effusion). Since incorrect attributes will lead to imprecise radiology reports, strengthening the generation process with precise attribute modeling becomes necessary. Additionally, the temporal information contained in the historical records, which is crucial in evaluating a patient’s current condition (e.g., heart size is unchanged), has also been largely disregarded. To address these issues, we propose RECAP, which generates precise and accurate radiology reports via dynamic disease progression reasoning. Specifically, RECAP first predicts the observations and progressions (i.e., spatiotemporal information) given two consecutive radiographs. It then combines the historical records, spatiotemporal information, and radiographs for report generation, where a disease progression graph and dynamic progression reasoning mechanism are devised to accurately select the attributes of each observation and progression. Extensive experiments on two publicly available datasets demonstrate the effectiveness of our model.

RECAP: Towards Precise Radiology Report Generation via Dynamic Disease Progression Reasoning

The recent success of general-domain large language models (LLMs) has significantly changed the natural language processing paradigm towards a unified foundation model across domains and applications. In this paper, we focus on assessing the performance of GPT-4, the most capable LLM so far, on the text-based applications for radiology reports, comparing against state-of-the-art (SOTA) radiology-specific models. Exploring various prompting strategies, we evaluated GPT-4 on a diverse range of common radiology tasks and we found GPT-4 either outperforms or is on par with current SOTA radiology models. With zero-shot prompting, GPT-4 already obtains substantial gains 
($\approx$ 10\% absolute improvement) over radiology models in temporal sentence similarity classification (accuracy) and natural language inference ($F_1$). For tasks that require learning dataset-specific style or schema (e.g. findings summarisation), GPT-4 improves with example-based prompting and matches supervised SOTA. Our extensive error analysis with a board-certified radiologist shows GPT-4 has a sufficient level of radiology knowledge with only occasional errors in complex context that require nuanced domain knowledge. For findings summarisation, GPT-4 outputs are found to be overall comparable with existing manually-written impressions.

Exploring the Boundaries of GPT-4 in Radiology | VIDEO

Automatically generated reports from medical images promise to improve the workflow of radiologists. Existing methods consider an image-to-report modeling task by directly generating a fully-fledged report from an image. However, this conflates the content of the report (e.g., findings and their attributes) with its style (e.g., format and choice of words), which can lead to clinically inaccurate reports. To address this, we propose a two-step approach for radiology report generation. First, we extract the content from an image; then, we verbalize the extracted content into a report that matches the style of a specific radiologist. For this, we leverage RadGraph---a graph representation of reports---together with large language models (LLMs). In our quantitative evaluations, we find that our approach leads to beneficial performance. Our human evaluation with clinical raters highlights that the AI-generated reports are indistinguishably tailored to the style of individual radiologist despite leveraging only a few examples as context.

Style-Aware Radiology Report Generation with RadGraph and Few-Shot Prompting

OBJECTIVE
Despite rising popularity and performance, studies evaluating the use of LLMs for clinical decision support are lacking. Here, we evaluate ChatGPT 3.5 and 4’s capacity for clinical decision support in radiology via the identification of appropriate imaging services for two important clinical presentations: breast cancer screening and breast pain.
METHODS
We compared ChatGPT's responses to the American College of Radiology (ACR) Appropriateness Criteria for breast pain and breast cancer screening. Our prompt formats included an open-ended (OE) and a select all that apply (SATA) format. Scoring criteria evaluated whether proposed imaging modalities were in accordance with ACR guidelines. Three replicate entries were conducted for each prompt, and the average of these was used to determine final scores.
RESULTS
Both ChatGPT 3.5 and 4 achieved an average OE score of 1.830 (out of 2) for breast cancer screening prompts. ChatGPT 3.5 achieved a SATA average percentage correct of 88.9%, compared to ChatGPT 4’s average percentage correct of 98.4% for breast cancer screening prompts. For breast pain, ChatGPT 3.5 achieved an average OE score of 1.125 (out of 2) and a SATA average percentage correct of 58.3%, as compared to an average OE score of 1.666 (out of 2) and a SATA average percentage correct of 77.7%. 
DISCUSSION
Our results demonstrate the eventual feasibility of using LLMs like ChatGPT for radiologic decision making, with the potential to improve clinical workflow and responsible use of radiology services. More use cases and greater accuracy are necessary to evaluate and implement such tools.


Evaluating GPT as an Adjunct for Radiologic Decision Making: GPT-4 Versus GPT-3.5 in a Breast Imaging Pilot

Background:
Magnetic Resonance Imaging (MRI) has been a large component of the evaluation of breast lesions in high-risk women for the last decade. Many of these imaging tools have been developed to assist physicians in risk stratifying patients. Specifically, Time Activity Curves (TAC) have been utilized to incorporate contrast uptake patterns into decision making. In the years since its inception, new tools have been developed to more efficiently and accurately classify different TACs, such as DynaCAD. Previous studies on the use of TAC curves have indicated their accuracy is not sufficient to be utilized without morphological and historical data being incorporated, but it is unclear how the improvements noted in these tools affects these results. Therefore, we re-evaluate TACs within a broad patient population and treatment landscape.
Methods:
We performed a retrospective chart review of 48 patients who received both an MRI with DynaCAD TAC analysis followed by a diagnostic biopsy. Of the 48 patients evaluated, 18 were malignant and 30 were benign. We performed statistical analysis for sensitivity, specificity, negative predictive value (NPV) and positive predictive value (PPV) on this set of patient results, stratifying by the individual kinetic classification as well as the generally accepted associations of each type with malignancy (Type 1 classically benign, Type 2 and 3, concerning for malignancy). We performed a literature search for previously reported TAC statistics.
Results:
From the 48 patients evaluated, washout curves utilized alone as a marker of malignancy status resulted in a sensitivity of 64.3%, specificity of 73.5% and PPV of 50% and NPV of 83%. When stratified by individual TAC curve type, persistent was calculated to have 50% and 81.1% sensitivity and specificity, respectively. Plateau was calculated to have a 72% and 94% sensitivity and specificity, respectively. Washout was calculated to have a 78.3% and 76.9% sensitivity and specificity, respectively.
Conclusion:
For years, TAC curves have been utilized to enhance radiologists’ ability to classify breast lesions as likely benign or malignant in high-risk patients. In our study we evaluated the accuracy of TACs with the new improvements provided by more advanced commercial tools. We have found TAC curves have improved in their accuracy with the rapid improvement of analysis tools. However, their primary benefit lies in being an ancillary tool for determining benignlesions along with morphological data and they still struggle in accurately predicting malignancy on biopsy. The study is limited due to its lower total patient population.

Revisiting Time Activity Curves in Breast MRI: Has it kept up with the times?

Purpose: MRI may be useful to identify women with cervix cancer at high risk of disease progression to test strategies of treatment intensification. The purpose of this study was to determine the value of a machine-learning model built on pre-treatment MRI for prediction of risk of progression after radiation therapy. 

Methods/Materials: Magnetic Resonance Imaging (MRI) Data for 54 women with cervix cancer was collected from The Cancer Genome Atlas Cervical Squamous Cell Carcinoma and Endocervical Adenocarcinoma Collection (TCGA-CESC) on the Cancer Imaging Archive (TCIA), which reported clinical, treatment, and imaging data from a single institution. 27 patients that had received radiation for cervix cancer were selected for input into a custom 3-D Residual Neural Network (ResNet) model with added custom layers specific to DICOM data in tensorflow python package. One T2 MRI per patient was used to predict progression free survival after radiation treatment, where patients were predicted to be “high risk” or “low risk” of progression as the output of the model. All slices of the T2 MRI were used. Model was validated using 5-fold cross validation; where 80% of the data was used to train each fold and 20% was used for validation. Final model statistical significance was confirmed through shuffle test at the p < 0.01 level. The clinical outcomes of patients predicted to be low risk or high risk by the machine learning model are reported.

Results: Median follow-up of patients was 29 months (range 3-64). Mean age was 51 years (range 29-79). 10 women were treated with radiation alone and 17 with chemo-radiation. 5 women received surgery in addition to radiation or chemoradiation. 21 patients received brachytherapy. Staging was IB in 9 women, IIA in 2, IIB in 9, IIIA in 2, IIIB in 2, IV in 3. 20 patients had squamous cell carcinoma and 7 patients had adenocarcinoma. 7 patients were predicted to be high risk by the model, and of these 7 developed progression during follow up. None of the 20 patients predicted to be low risk developed progression. Among all patients in the study, the 2 year progression free survival (PFS) was 82.0%. Among patients predicted to be low risk by model, the 2 year PFS was 100%, and for patients predicted to be high risk by model, the 2 year PFS was 43%. Of patients with progression, 3 developed local recurrence and 4 developed distant progression. ResNet model achieved cross-validated accuracy of 92% for prediction of progression free survival. Model was significant at p < 0.01 by shuffle test.

Conclusion: A 3-D ResNet machine-learning model built upon pre-treatment MRI data for cervical cancer was able to accurately predict clinical outcomes following radiation therapy in a single institution dataset. Future work to confirm generalizability should focus on validation on a larger clinical dataset.


Machine-Learning Model of Pre-Treatment MRI Accurately Predicts Progression Following Cervix Cancer Radiotherapy

Background:
Breast cancer is the most commonly diagnosed cancer type, accounting for about 2.3 million new cases and 685,000 deaths worldwide in 2020. In the US, breast cancer screenings are routinely performed in women starting from the age of 40. One major challenge with breast screening is the high rate of false positives, with over 50% of women experiencing at least one false positive result over 10 years of annual screening on plain film mammography, the most common breast cancer screening modality. Deep learning methods have been shown to be effective in reducing false positive rates that are associated with unnecessary costs, medical procedures, and emotional burden on patients, as well as improving the accuracy of cancer detection in screening mammograms — hence, this study investigates the use of deep learning methods for breast cancer detection. Our aim was to investigate and evaluate the performance of transfer learning and deep learning models in breast cancer detection.

Methods:
We trained and evaluated our deep learning model using multi-center breast screening data provided by Kaggle in the RSNA Screening Mammography Breast Cancer Detection competition. Our model uses mammogram images as input to fine-tune a pre-trained ResNet-50 deep-learning model via transfer learning and outputs the probability of cancer in each image. The model performance is evaluated using a held-out test-set (train:validate:test 70:10:20) to compute the sensitivity, specificity, and probabilistic F1-score achieved.

Results:
Overall we achieved a classification sensitivity of 1.00 and a specificity of 1.00 (ROC-AUC=1.00) on our held-out test set, with a probabilistic F1 score of 0.795 on our held-out test set. 

Conclusion:
Our model detected breast cancer with 100% sensitivity and specificity in the held-out test set. Future work may focus on validating algorithm performance in a larger out-of-distribution patient cohort. Application of deep learning models to breast cancer detection in mammograms may lead to improved clinical efficiency, diagnostic accuracy, and reduction in costs associated with the high false positive burden in breast cancer screening.


Deep Transfer Learning for Breast Cancer Detection Using 2D Screening Mammography

Background: The EMBARK trial is a phase 3, randomized, study evaluating use of enzalutamide with or without concomitant leuprolide in non-metastatic castrate sensitive prostate cancer patients. Trial enrollment relied upon conventional imaging which under detects metastatic disease in comparison to PSMA-PET imaging. The aim of this post-hoc retrospective analysis was to describe the PSMA-PET findings and survival outcomes in a patient population representative of the EMBARK trial. 

Methods: This was a post-hoc, retrospective analysis of 5 prospective studies of PSMA-PET conducted at UCLA from 2016 to 2021 that included patients with recurrent PCa. Patients were included if they met eligibility EMBARK trial enrollment at the time of their PSMA scan as per the following: Rising PSA above 1.0ng/mL (post RP and SRT) or 2.0ng/mL >nadir (post-RT), PSA doubling time ≤9 months, Serum testosterone ≥150 ng/dL. PSMA PET findings (PROMISE miTNM stage, number of lesions) and survival outcomes were collected from clinical reports. Descriptive statistics (median, range) and Kaplan-Meier modeling (median, Confidence intervals) were used.

Results: From 2002 patients screened, 146 patients were included in the analysis. Median pre-scan PSA levels and PSA doubling time in patients who underwent RP (n=85), dRT (n=22), post-RP SRT (n=39) and overall (n=146) were 2.4 ng/mL (1-72.1), 8.3 ng/mL (2.37-202), 2.58 ng/ml (1.1-32.3), 2.7 ng/mL (1-202), and 3.5 months (1-9), 3.3 months (1-9), 4.4 months (1-9), 3.7 (1-9) months, respectively. PSMA-PET was positive in 80%, 41%, 85%, 83% of RP, dRT, RP + SRT and overall populations, respectively. Poly-metastatic disease (≥5 lesions) was found in 19%, 32%, 21% and 21% of cases, respectively. PSMA-PET detected disease only localized to the prostate fossa (miT+N0M0) in 7%, 32%, 0% and 9%. PSMA-PET detected any distant metastatic disease (miTxNxM1) in 33%, 41% 56% and 40% respectively. PSMA-PET detected metastatic nodal disease (N1 and/or M1a) in 58% 32% 74% and 39% and osseous disease (M1b) in 19% 23% 31% and 23% cases. Twenty-eight patients received androgen receptor-axis therapy (ARSi) or androgen deprivation therapy (ADT) following PSMA-PET had available outcome data. Recurrence free survival in ARSi compared to the ADT cohort was 53 months vs 22 months (p=0.008).
Conclusion: In this “EMBARK-like” cohort of patients with high-risk nmCSPCa and PSA DT<9 months, PSMA-PET detected M1 disease in 40% of patients, over one-fifth of cases presented with 5 or more lesions and those receiving ARSi had significantly improved recurrence free survival. Further studies are needed to assess its independent prognostic value and use for treatment guidance.


PSMA PET findings in an “EMBARK-like” cohort of patients with high-risk non-metastatic hormone-sensitive prostate cancer: A single center post-hoc retrospective analysis

Background: Prostate cancer (PCa) is the second most commonly diagnosed malignancy and the second leading cause of cancer death in men in the United States. Traditional screening for PCa has included monitoring of prostate specific antigen (PSA) levels and systematic transrectal ultrasound-guided (TRUS) prostate biopsy. Recommendations for PCa screening are evolving with recent advances in multi-parametric MRI (mp-MRI) of the prostate. However, negative findings on mp-MRI of the prostate in patients with suspicion of PCa can lead to uncertainties regarding clinical management. This study therefore aims to evaluate the detection rate of PCa by TRUS biopsy in patients with negative mp-MRI.

Methods: The records of 253 patients without evidence of PCa on mp-MRI performed at the Virginia Commonwealth University Health System between 2018 and 2023 who also underwent TRUS prostate biopsies were retrospectively reviewed. Clinical information including age at mp- MRI, PSA levels, and prostate volume, as well as surgical pathology results, namely presence of malignancy, Gleason score (GS), and percentage of positive cores on biopsy, were recorded for these patients. Ultimately, the incidence of clinically significant PCa (GS ≥ 7) and clinically insignificant PCa (GS 6) detected on TRUS-guided biopsy was determined.

Results: Of 253 patients with negative mp-MRI findings, 126 (49.8%) had PCa at TRUS biopsy. Among these 126 patients, 37 had clinically significant PCa with GS 7-9 (14.6% of patients with negative mp-MRI, 29.4% of those with biopsy-proven PCa), and 89 had clinically insignificant PCa with GS 6 (35.2% of patients with negative mp-MRI, 70.6% of those with biopsy-proven PCa). Of the patients with clinically significant PCa, 89.2% (33 patients) had PCa up to GS 7, 5.41% (2 patients) had PCa up to GS 8, and 5.41% (2 patients) had PCa up to GS 9 on TRUS-
guided biopsy. The remaining 127 patients (50.2% of total patients with negative mp-MRI) had no evidence of PCa on TRUS biopsy. There was no statistically significant difference in age, PSA level, or prostate volume between patients with PCa and without PCa at biopsy.

Conclusion: This retrospective study confirms that a percentage of patients without evidence of PCa at mp-MRI of the prostate may still have clinically significant PCa and may benefit from further screening, including TRUS biopsy. Further research may be of benefit to determine optimal screening for PCa.

Detection Rate of Prostate Cancer at Systematic Prostate Biopsy in Patients with Negative mp-MRI of the Prostate: A Retrospective Study

**Scoring is only possible in the semifinals hall**

**Abstract:**



Abbreviated Breast MRI Protocol: A Feasible Alternative for Breast Screening

T2 White Matter Hyperintensity Volumes Are Correlated With Poststroke Neuropsychological Deficits

Background
Renal cell carcinoma (RCC) is a highly heterogenous cancer that typically relies on biopsy to identify the specific tumor subtype and identify high-grade histology. However, accuracy and reliability of biopsy is sometimes limited by tumor heterogeneity. The use of 3D printing to improve the accuracy of this process has been previously described for prostate and renal cancers. Preoperative imaging is used to 3D print “negative” molds matching the contour of the surgically removed tumor and provide guiding lines along which to cut precise slices that correspond with imaging slices. Our objective was to design a novel technique combining negative molds with additional 3D-printed overlays, or “positive” molds that would fit onto the pathology slices and have cutouts to denote regions of interest previously identified on imaging, further facilitating the sectioning process and correlating pathology findings to imaging in RCC.

Methods
This was a single-center, IRB-approved, prospective study in patients with RCC who planned to undergo partial or radical nephrectomy. Preoperative images were transferred to a viewing software (3D Slicer), and their data was used to generate “positive” and “negative” 3D -printed molds. Following nephrectomy, kidneys were sectioned by a study pathologist and study radiologist using the 3D-printed molds. Pathology slides were then directly compared to CT by a radiologist using 3D Slicer and Docker to determine degree of pathology-to-radiology correlation. 

Results
11 total 3D-printed molds were constructed. Mold failure was defined by cases where 3D-printed molds could not be used throughout the entire sectioning process, or where lesions on pathology did not align with preoperative imaging. Mold success was defined by cases with complete mold-assisted sectioning and correlation of pathology to radiology images. The first 5 molds produced failed, either due to tumor growth between imaging and sectioning (n=3) or perirenal fat (n=2). Improvements in mold design using a bivalving technique were achieved following discussion between the study urologist, pathologist, and radiologist, and the subsequent 6 molds produced were successful.

Conclusion
In this study, we aimed to develop a novel technique for designing patient-specific renal molds to improve the diagnostic accuracy of tumor grading in renal cell carcinoma (RCC), which we achieved following multidisciplinary discussion and process revision. This is the first study to describe such an approach using negative and positive 3D-printed molds, and further research is warranted to validate the efficacy and efficiency of our proposed methodology in order to better standardize tumor sectioning and advance the diagnostic process for RCC.


Underline digital video library

RECAP: Towards Precise Radiology Report Generation via Dynamic Disease Progression Reasoning

Exploring the Boundaries of GPT-4 in Radiology | VIDEO

Style-Aware Radiology Report Generation with RadGraph and Few-Shot Prompting

Evaluating GPT as an Adjunct for Radiologic Decision Making: GPT-4 Versus GPT-3.5 in a Breast Imaging Pilot

Revisiting Time Activity Curves in Breast MRI: Has it kept up with the times?

Machine-Learning Model of Pre-Treatment MRI Accurately Predicts Progression Following Cervix Cancer Radiotherapy

Deep Transfer Learning for Breast Cancer Detection Using 2D Screening Mammography

PSMA PET findings in an “EMBARK-like” cohort of patients with high-risk non-metastatic hormone-sensitive prostate cancer: A single center post-hoc retrospective analysis

Detection Rate of Prostate Cancer at Systematic Prostate Biopsy in Patients with Negative mp-MRI of the Prostate: A Retrospective Study

Abbreviated Breast MRI Protocol: A Feasible Alternative for Breast Screening

T2 White Matter Hyperintensity Volumes Are Correlated With Poststroke Neuropsychological Deficits

3D Printed Molds for Pathology Sectioning of Renal Masses: Enabling Radiology-Pathology Correlation

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES