Mammography Breast Cancer Screening Triage Using Deep Learning: A UK Retrospective Study.

Background Breast screening enables early detection of cancers; however, most women have normal mammograms, resulting in repetitive and resource-intensive reading tasks. Purpose To investigate if deep learning (DL) algorithms can be used to triage mammograms by identifying normal results to reduce workload or flag cancers that may be overlooked. Materials and Methods In this retrospective study, three commercial DL algorithms were investigated using consecutive mammograms from two UK Breast Screening Program sites from January 2015 to December 2017 and January 2017 to December 2018 on devices from two mammography vendors. Normal mammograms with a 3-year follow-up and histopathologically proven cancer detected at screening, the subsequent round, or in the 3-year interval were included. Two algorithm thresholds were set: in scenario A, 99.0% sensitivity for rule-out triage to a lone reader, and in scenario B, approximately 1.0% additional recall providing a rule-in triage for further assessment. Both thresholds were then applied to the screening workflow in scenario C. The sensitivity and specificity were used to assess the overall predictive performance of each DL algorithm. Results The data set comprised 78 849 patients (median age, 59 years [IQR, 53-63 years]) and 887 screening-detected, 439 interval, and 688 subsequent screening round-detected cancers. In scenario A (rule-out triage), models DL-1, DL-2, and DL-3 triaged 35.0% (27 565 of 78 849), 53.2% (41 937 of 78 849), and 55.6% (43 869 of 78 849) of mammograms, respectively, with 0.0% (0 of 887) to 0.1% (one of 887) of screening-detected cancers undetected. In scenario B, DL algorithms triaged in 4.6% (20 of 439) to 8.2% (36 of 439) of interval and 5.2% (36 of 688) to 6.1% (42 of 688) of subsequent-round cancers when applied after the routine double-reading workflow. Combining both approaches in scenario C resulted in an overall noninferior specificity (difference, -0.9%; P < .001) and superior sensitivity (difference, 2.7%; P < .001) for the adaptive workflow compared with routine double reading for all three algorithms. Conclusion Rule-out and rule-in DL-adapted triage workflows can improve the efficiency and efficacy of mammography breast cancer screening. © RSNA, 2023 Supplemental material is available for this article. See also the editorial by Nishikawa and Lu in this issue.

Constrained surface controllers for three-dimensional image data reformatting.

This study did not require ethical approval in the United Kingdom. The aim of this work was to create two controllers for navigating a two-dimensional image plane through a volumetric data set, providing two important features of the ultrasonographic paradigm: orientation matching of the navigation device and the desired image plane in the three-dimensional (3D) data and a constraining surface to provide a nonvisual reference for the image plane location in the 3D data. The first constrained surface controller (CSC) uses a planar constraining surface, while the second CSC uses a hemispheric constraining surface. Ten radiologists were asked to obtain specific image reformations by using both controllers and a commercially available medical imaging workstation. The time taken to perform each reformatting task was recorded. The users were also asked structured questions comparing the utility of both methods. There was a significant reduction in the time taken to perform the specified reformatting tasks by using the simpler planar controller as compared with a standard workstation, whereas there was no significant difference for the more complex hemispheric controller. The majority of users reported that both controllers allowed them to concentrate entirely on the reformatting task and the related image rather than being distracted by the need for interaction with the workstation interface. In conclusion, the CSCs provide an intuitive paradigm for interactive reformatting of volumetric data.

An investigation into the effects of temporal resolution on hepatic dynamic contrast-enhanced MRI in volunteers and in patients with hepatocellular carcinoma.

This study investigated the effect of temporal resolution on the dual-input pharmacokinetic (PK) modelling of dynamic contrast-enhanced MRI (DCE-MRI) data from normal volunteer livers and from patients with hepatocellular carcinoma. Eleven volunteers and five patients were examined at 3 T. Two sections, one optimized for the vascular input functions (VIF) and one for the tissue, were imaged within a single heart-beat (HB) using a saturation-recovery fast gradient echo sequence. The data was analysed using a dual-input single-compartment PK model. The VIFs and/or uptake curves were then temporally sub-sampled (at interval ▵t = [2-20] s) before being subject to the same PK analysis. Statistical comparisons of tumour and normal tissue PK parameter values using a 5% significance level gave rise to the same study results when temporally sub-sampling the VIFs to HB < ▵t <4 s. However, sub-sampling to ▵t > 4 s did adversely affect the statistical comparisons. Temporal sub-sampling of just the liver/tumour tissue uptake curves at ▵t ≤ 20 s, whilst using high temporal resolution VIFs, did not substantially affect PK parameter statistical comparisons. In conclusion, there is no practical advantage to be gained from acquiring very high temporal resolution hepatic DCE-MRI data. Instead the high temporal resolution could be usefully traded for increased spatial resolution or SNR.

Slice offset frequency and shim adjustment for interactive steady-state free-precession (SSFP) imaging.

PURPOSE: To devise a method allowing real-time optimization of center frequency (CF) and shim for an interactive steady-state free-precession (SSFP) sequence by reformatting a previously acquired field map in the same orientation as the interactive acquisition. MATERIALS AND METHODS: Field maps were acquired in a rectangular parallel-piped phantom and a normal volunteer. An SSFP sequence was modified to communicate the current slice offset and rotation to an external program that reformatted the field map into the same plane, calculated the CF and shim offsets, and passed them back to the sequence. CF offsets as a function of position for the phantom were compared with the scanner prescan-determined offset. RESULTS: In the phantom, the CF measurements agreed with the scanner-determined offsets. Bland-Altman analysis showed a bias of -14 Hz (field map - prescan) and limits of agreement of -28 to 0 Hz. In the volunteer there was a qualitative improvement in image quality when using the optimized center frequencies and shims. CONCLUSION: The proposed method demonstrates how CF and shim can be optimized for any interactively positioned slice, resulting in reduced off-resonance artifacts.

MR-guided direct arthrography of the hip.

PURPOSE: To develop an MR-guided technique for direct magnetic resonance arthrography (DMRA) of the hip using a conventional 1.5 Tesla (T) MRI scanner. MATERIALS AND METHODS: Interactive versions of standard single-shot fast spin echo, coherent gradient echo, and fully balanced steady-state free precession sequences were developed, together with the ability to operate the scanner from within the magnet room. Initial optimization was performed using ex vivo animal joints. After informed consent, five patients underwent DMRA using dilute gadolinium as the contrast medium. RESULTS: The procedure was successful in all five patients and diagnostic quality MR arthrography studies were obtained in all cases. The time taken from the initial placement of skin marker to the introduction of contrast medium reduced with successive patients: 26, 22, and 19 min with the final two studies taking 13 min each. CONCLUSION: This work demonstrates the feasibility of performing DMRA using MR guidance, avoiding the need for X-ray fluoroscopy facilities but exploiting the available imaging performance of a high-field closed-bore MR system. The use of an in room control and display system together with interactive fluoroscopic imaging sequences simplifies the positioning of the needle and allows rapid confirmation of an intra-articular injection.

Interactive magnetic resonance cholangiography (MRC) with adaptive averaging.

PURPOSE: To implement and evaluate a technique for adaptively averaging 2D magnetic resonance cholangiography (MRC) images obtained using an interactive imaging system with a view to improving image quality at reduced fields of view (FOVs). MATERIALS AND METHODS: Images were obtained using an interactive implementation of a single-shot half-Fourier rapid acquisition with relaxation enhancement (RARE) technique. Software was developed for adaptively averaging images, and an evaluation was performed in a phantom and a cohort of 10 patients referred for standard MRC. Adaptively averaged and standard single-shot MRC images were evaluated with respect to their ability to demonstrate the common bile duct and main left and right intrahepatic duct branches. RESULTS: In all patient studies there was no difference in the ability of either the adaptive technique or the standard single-shot method to demonstrate the common bile duct and the main left and right intrahepatic duct branches. However, in seven of the 10 patient studies the adaptive technique provided better visualization of the peripheral bile duct system (P = 0.035; sign test). There was no difference in the diagnostic confidence of the two techniques (P = 0.32, Wilcoxon signed-rank test). CONCLUSION: Adaptive averaging of MRC images obtained using an interactive imaging paradigm significantly improves visualization of peripheral intrahepatic ducts.