Deep learning in breast radiology: current progress and future directions.

This review provides an overview of current applications of deep learning methods within breast radiology. The diagnostic capabilities of deep learning in breast radiology continue to improve, giving rise to the prospect that these methods may be integrated not only into detection and classification of breast lesions, but also into areas such as risk estimation and prediction of tumor responses to therapy. Remaining challenges include limited availability of high-quality data with expert annotations and ground truth determinations, the need for further validation of initial results, and unresolved medicolegal considerations. KEY POINTS: • Deep learning (DL) continues to push the boundaries of what can be accomplished by artificial intelligence (AI) in breast imaging with distinct advantages over conventional computer-aided detection. • DL-based AI has the potential to augment the capabilities of breast radiologists by improving diagnostic accuracy, increasing efficiency, and supporting clinical decision-making through prediction of prognosis and therapeutic response. • Remaining challenges to DL implementation include a paucity of prospective data on DL utilization and yet unresolved medicolegal questions regarding increasing AI utilization.

The Interplay between Molten Globules and Heme Disassociation Defines Human Hemoglobin Disassembly.

Hemoglobin functions as a tetrameric oxygen transport protein, with each subunit containing a heme cofactor. Its denaturation, either in vivo or in vitro, involves autoxidation to methemoglobin, followed by cofactor loss and globin unfolding. We have proposed a global disassembly scheme for human methemoglobin, linking hemin (ferric protoporphyrin IX) disassociation and apoprotein unfolding pathways. The model is based on the evaluation of circular dichroism and visible absorbance measurements of guanidine-hydrochloride-induced disassembly of methemoglobin and previous measurements of apohemoglobin unfolding. The populations of holointermediates and equilibrium disassembly parameters were estimated quantitatively for adult and fetal hemoglobins. The key stages are characterized by hexacoordinated hemichrome intermediates, which are important for preventing hemin disassociation from partially unfolded, molten globular species during early disassembly and late-stage assembly events. Both unfolding experiments and independent small angle x-ray scattering measurements demonstrate that heme disassociation leads to the loss of tetrameric structural integrity. Our model predicts that after autoxidation, dimeric and monomeric hemichrome intermediates occur along the disassembly pathway inside red cells, where the hemoglobin concentration is very high. This prediction suggests why misassembled hemoglobins often get trapped as hemichromes that accumulate into insoluble Heinz bodies in the red cells of patients with unstable hemoglobinopathies. These Heinz bodies become deposited on the cell membranes and can lead to hemolysis. Alternatively, when acellular hemoglobin is diluted into blood plasma after red cell lysis, the disassembly pathway appears to be dominated by early hemin disassociation events, which leads to the generation of higher fractions of unfolded apo subunits and free hemin, which are known to damage the integrity of blood vessel walls. Thus, our model provides explanations of the pathophysiology of hemoglobinopathies and other disease states associated with unstable globins and red cell lysis and also insights into the factors governing hemoglobin assembly during erythropoiesis.

THE TEXAS TACO TECHNIQUE FOR INTERNAL LIMITING MEMBRANE FLAP IN LARGE FULL-THICKNESS MACULAR HOLES: A Short-Term Pilot Study.

PURPOSE: To determine the efficacy of a novel application of a surgical internal limiting membrane flap technique that requires no additional surgical adjuvants in closure of large full-thickness macular holes (FTMHs). METHODS: The electronic medical records of patients (n = 8) with large (>400 µm) FTMHs repaired with the "Texas Taco" technique were retrospectively reviewed. RESULTS: Operated patients had a mean age of 63.8 ± 19.2 (range, 19-80) years. There were five (62.5%) phakic and three (37.5%) pseudophakic eyes preoperatively. Mean follow-up time was 9.1 ± 4.7 (1.5-14.5) months. Across all patients, mean FTMH diameter at the shortest and greatest widths were 529 ± 101 (404-661) and 1,189 ± 290 (829-1,656) µm, respectively. Mean best-corrected logarithm of the minimum angle of resolution visual acuity was 1.3 ± 0.23 preoperatively (approximately Snellen acuity 20/400) and 0.66 ± 0.40 postoperatively (approximately Snellen acuity 20/100) (P < 0.001). All FTMHs remained closed at all postoperative visits. CONCLUSION: The Texas Taco technique provided anatomical and functional improvement in challenging cases of large FTMHs without the need of additional surgical adjuvants.

Targeted Retinal Photocoagulation for Diabetic Macular Edema with Peripheral Retinal Nonperfusion: Three-Year Randomized DAVE Trial.

PURPOSE: To evaluate the effect of targeted retinal photocoagulation (TRP) on visual and anatomic outcomes and treatment burden in eyes with diabetic macular edema (DME). DESIGN: Phase I/II prospective, randomized, controlled clinical trial. PARTICIPANTS: Forty eyes of 29 patients with center-involved macular edema secondary to diabetes mellitus. METHODS: Eyes with center-involved DME and Early Treatment Diabetic Retinopathy Study (ETDRS) best-corrected visual acuity (BCVA) between 20/32 and 20/320 (Snellen equivalent) were randomized 1:1 to monotherapy with 0.3 mg ranibizumab (Lucentis, Genentech, South San Francisco, CA) or combination therapy with 0.3 mg ranibizumab and TRP guided by widefield fluorescein angiography. All eyes received 4 monthly ranibizumab injections followed by monthly examinations and pro re nata (PRN) re-treatment through 36 months. Targeted retinal photocoagulation was administered outside the macula to areas of retinal capillary nonperfusion plus a 1-disc area margin in the combination therapy arm at week 1, with re-treatment at months 6, 18, and 25, if indicated. MAIN OUTCOME MEASURES: Mean change in ETDRS BCVA from baseline and number of intravitreal injections administered. RESULTS: At baseline, mean age was 55 years, mean BCVA was 20/63 (Snellen equivalent), and mean central retinal subfield thickness (CRT) was 530 µm. Thirty-four eyes (85%) completed month 36, at which point mean BCVA improved 13.9 and 8.2 letters (P = 0.20) and mean CRT improved 302 and 152 µm (P = 0.03) in the monotherapy and combination therapy arms, respectively. The mean number of injections administered through month 36 was 24.4 (range, 10-34) and 27.1 (range, 12-36), with 73% (362/496) and 80% (433/538) of PRN injections administered (P = 0.004) in the monotherapy and combination therapy arms, respectively. Goldmann visual field isopter III-4e area decreased by 2% and 18% in the monotherapy and combination therapy arms, respectively (P = 0.30). CONCLUSIONS: In this 3-year randomized trial of 40 eyes with DME, there was no evidence that combination therapy with ranibizumab and TRP improved visual outcomes or reduced treatment burden compared with ranibizumab alone.

Mechanism of Human Apohemoglobin Unfolding.

Removal of heme from human hemoglobin (Hb) results in formation of an apoglobin heterodimer. Titration of this apodimer with guanidine hydrochloride (GdnHCl) leads to biphasic unfolding curves indicating two distinct steps. Initially, the heme pocket unfolds and generates a dimeric intermediate in which ∼50% of the original helicity is lost, but the α1ß1 interface is still intact. At higher GdnHCl concentrations, this intermediate dissociates into unfolded monomers. This structural interpretation was verified by comparing GdnHCl titrations for adult human hemoglobin A (HbA), recombinant fetal human hemoglobin (HbF), recombinant Hb cross-linked with a single glycine linker between the α chains, and recombinant Hbs with apolar heme pocket mutations that markedly stabilize native conformations in both subunits. The first phase of apoHb unfolding is independent of protein concentration, little affected by genetic cross-linking, but significantly shifted toward higher GdnHCl concentrations by the stabilizing distal pocket mutations. The second phase depends on protein concentration and is shifted to higher GdnHCl concentrations by genetic cross-linking. This model for apoHb unfolding allowed us to quantitate subtle differences in stability between apoHbA and apoHbF, which suggest that the ß and γ heme pockets have similar stabilities, whereas the α1γ1 interface is more resistant to dissociation than the α1ß1 interface.

Normal- and Low-Luminance Automated Quantitative Contrast Sensitivity Assessment in Eyes With Age-Related Macular Degeneration.

PURPOSE: To assess the effectiveness of an active learning approach to measuring the contrast sensitivity function (CSF) in patients with various degrees of dry age-related macular degeneration (AMD) under multiple luminance conditions. DESIGN: Cross-sectional study. METHODS: Patients with AMD (26 intermediate AMD, 19 AMD with subretinal drusenoid deposits [SDD], 20 geographic atrophy [GA]) and 23 age-matched controls were tested with the Manifold Contrast Vision Meter (Adaptive Sensory Technology) and the qCSF algorithm, which applies active learning to estimate a model of the CSF's global shape. Testing was performed under conditions of standard and low luminance. For each AMD severity, the area under log CSF (AULCSF) and contrast sensitivities at individual spatial frequencies were calculated for analysis. Low-luminance deficits (LLDs) for visual acuity (VA) and AULCSF were calculated as the difference between standard and low luminance values. RESULTS: Progressive decreases in AULCSF were observed as disease severity increased. For standard luminance, pairwise comparisons revealed significant differences between control/intermediate AMD (P < .0005), control/SDD (P < .0005), control/GA (P < .0005), and intermediate AMD/GA (P < .005). Similarly, for low luminance, pairwise comparisons revealed significant differences between the controls and each disease group (all P < .0005), in addition to significant differences between intermediate AMD/SDD (P < .005), and intermediate AMD/GA (P < .005). No correlations were found between LLD VA and LLD AULCSF in any AMD groups. CONCLUSIONS: Contrast sensitivity measured via qCSF under both standard- and low-luminance conditions correlates with advancing stages of dry AMD. The interaction between luminance and contrast sensitivity appears to reflect a different aspect of visual function than the interaction between luminance and VA.

The Relationship Between Central Drusen Volume and Low-Luminance Deficit in Age-Related Macular Degeneration.

Purpose: To determine the relationship between central drusen volume and low-luminance deficit (LLD) in visual acuity (VA) in patients with intermediate age-related macular degeneration (AMD). Methods: In this cross-sectional study, 42 patients with intermediate AMD underwent testing for VA and low-luminance VA (LLVA), as well as spectral-domain optical coherence tomography. LLD was calculated as the difference between VA and LLVA. Central drusen volume was measured in the central 3 mm of the macula, defined as the volume between the inner border of the retinal pigment epithelium and Bruch's membrane. Results: Mean ± standard deviation (SD) LLD was 0.32 ± 0.12 logMAR and mean ± SD drusen volume was 0.18 ± 0.09 mm3. No linear relationship was identified between central 3 mm drusen volume and LLD (P = 0.215). R2 for the bivariate linear model was 0.038 (95% confidence interval 0-0.222). Limitation of the analysis to drusen volumes measured in the central 1 mm of the macula did not impact results (R2 = 0.075), nor did incorporation of lens status into the model (R2 = 0.067) or censoring of patients with nonfoveal subretinal drusenoid deposits (R2 = 0.071). Conclusions: The amount of drusen within the central 3 mm of the macula does not appear to be related to LLD in intermediate AMD. These measures may be manifestations of different underlying pathophysiological mechanisms. Translational Relevance: Understanding relationships between markers for AMD progression may help guide development of improved clinical grading scales for AMD.

Longitudinal Quantification of Retinal Nonperfusion in the Macula of Eyes With Retinal Vein Occlusion Receiving Anti-VEGF Therapy: Secondary Analysis of the WAVE Randomized Trial.

BACKGROUND AND OBJECTIVE: Longitudinal quantification of retinal nonperfusion (RNP) in eyes with retinal vein occlusion (RVO) undergoing anti-vascular endothelial growth factor therapy. PATIENTS AND METHODS: Thirty eyes with ischemic RVO were randomized to ranibizumab (Lucentis; Genentech, South San Francisco, CA) (monotherapy) or ranibizumab plus peripheral laser (combination therapy) in a 12-month, prospective trial. RNP on fluorescein angiography was quantified within the macula through 12 months of follow-up. RESULTS: Baseline mean macular RNP areas were 5.04 mm2 and 8.30 mm2 in the monotherapy (n = 5) and combination therapy (n = 15) cohorts, respectively. Through month 12, mean macular RNP area increased 0.36 mm2 and 0.53 mm2 in the monotherapy and combination therapy cohorts, respectively (P = .77). Marked, progressive RNP was observed in three eyes (12%). No areas of reperfusion were detected in any eye. CONCLUSION: Among ischemic RVO eyes in WAVE, macular RNP was common at baseline and remained stable over time in most eyes, though marked RNP progression occurred in a minority of eyes. [Ophthalmic Surg Lasers Imaging Retina. 2018;49:258-264.].

Identification of factors to increase efficacy of telemedicine screening for diabetic retinopathy in endocrinology practices using the Intelligent Retinal Imaging System (IRIS) platform.

AIMS: Diabetic retinopathy (DR) and diabetic macular edema (DME) can be evaluated using telemedicine systems, such as the Intelligent Retinal Imaging Systems (IRIS), in patients with Diabetes Mellitus (DM). In an endocrinology-based population utilizing IRIS we determine prevalence rates of DR and DME, and identify associated epidemiologic correlations. METHODS: This is a multicenter, retrospective chart review using screening data from IRIS. Centers for Disease Control and Prevention (CDC) data on epidemiologic variables (by county) namely, prevalence of DM, incidence of DM, obesity, and time of physical inactivity, were compared against prevalence rates of DR found at screening. RESULTS: A total of 10,223 eyes of 5,242 patients with DM were imaged. DR and DME were noted in 1781 (33.98%) and 226 imaging studies (4.31%) respectively. The coefficient of determination was greatest for incidence of DM (R2 = 0.92), followed by DM prevalence (R2 = 0.79), obesity, (R2 = 0.67), and physical inactivity (R2 = 0.34). The presence of DR during screening varied significantly by county (p < 0.001). CONCLUSIONS: Screening in counties with a higher incidence of DM led to a higher prevalence of identified DR at time of screening. The current work suggests that telemedicine screening in areas known to have a higher incidence of DM may be worthwhile.

Predictors of Visual Acuity Outcomes Following Vitrectomy for Idiopathic Macular Hole.

BACKGROUND AND OBJECTIVE: To investigate predictors of visual outcomes in patients who underwent vitrectomy for full-thickness macular hole (FTMH) with at least 1 year of follow-up. PATIENTS AND METHODS: Retrospective, noncomparative, consecutive case series of 132 eyes of 122 patients who underwent surgical repair of idiopathic FTMH with at least 1 year of follow-up. Predictors of visual acuity (VA) outcomes were analyzed using linear regression. RESULTS: Mean follow-up time was 22.2 months. Twenty-three eyes (17.4%) had age-related macular degeneration (AMD), of which 17 (73.9%) cases were mild and nonexudative. At final follow-up, poor preoperative VA (P < .001), perioperative complications (P < .001), AMD (P < .001), and delay from preoperative evaluation to surgery (P = .037) were significant predictors of final VA. In multiple regression, these variables remained significant (P < .001, P = .011, P < .001, and P = .002, respectively). CONCLUSION: Poor preoperative VA, perioperative complications, AMD, and delay to surgery were significant predictors of final VA following FTMH repair. [Ophthalmic Surg Lasers Imaging Retina. 2018;49:566-570.].

Long-term outcomes with as-needed aflibercept in diabetic macular oedema: 2-year outcomes of the ENDURANCE extension study.

BACKGROUND/AIMS: To evaluate the efficacy and safety of individualised 2.0 mg intravitreal aflibercept retreatment for diabetic macular oedema (DME) through the fifth year of management. METHODS: This is a phase IV, 2-year, open-label extension study. Sixty patients completing the 3-year VISTA DME (Study of Intravitreal Aflibercept Injection in Patients With Diabetic Macular Edema) phase III trial enrolled in the ENDURANCE (Long-Term Efficacy and Safety of Intravitreal Aflibercept for the Treatment of DME in Subjects Who Completed the VISTA DME Trial) extension study. All patients received aflibercept in the presence of clinically relevant DME. Intervals between visits were prescribed according to disease activity. The main outcome measure was mean aflibercept injections given through 2 years. RESULTS: A mean of 7.7 aflibercept injections were administered through 2 years. Fifteen (25%) patients required no retreatment and 48% (n=29) of patients received five or fewer injections through 2 years. Among patients who received at least one aflibercept retreatment during ENDURANCE, the mean number of injections through 2 years was 9.5. The mean visual acuity and central retinal thickness gains achieved during VISTA DME were maintained and stable during ENDURANCE. The most notable safety signal was progression of diabetic retinopathy. Six (10%) patients converted from non-proliferative to proliferative diabetic retinopathy (PDR), and a total of eight patients experienced PDR events occurring at a mean of 387 days following the previous aflibercept treatment. CONCLUSION: The need for aflibercept retreatment was substantially reduced in the fourth and fifth years of aflibercept dosing for DME following initiation of therapy in the VISTA DME trial. While vision gains achieved during the 3-year VISTA DME trial were maintained through ENDURANCE with a reduced treatment burden, clinically relevant worsening of diabetic retinopathy was observed with progression to PDR in 10% of the eyes. TRIAL REGISTRATION NUMBER: NCT02299336.

Neovascular age-related macular degeneration management in the third year: final results from the TREX-AMD randomised trial.

BACKGROUND/AIMS: Prospectively evaluate outcomes in the third year of neovascular age-related macular degeneration (AMD) management using ranibizumab with continued treat and extend (TREX) dosing compared with monthly visits with retreatment upon evidence of exudative disease activity (PRN, pro re nata). METHODS: Subjects with treatment-naïve neovascular AMD were randomised 1:2 to Monthly or TREX and managed through 2 years. In the third year, subjects randomised to Monthly were managed PRN while subjects randomised to TREX were continued on TREX dosing or transitioned to PRN after achieving an interval of 12 weeks between visits. RESULTS: Sixty subjects enrolled and 46 (77%) completed month 36 (M36). Transition from Monthly to PRN was associated with a decline in best corrected visual acuity (BCVA) (+10.5 letters (month 24) to +5.4 (M36, p=0.09)); three (15%) subjects required no dosing during year 3, and 47% (114/243) of possible PRN injections were delivered, yielding a mean of 6.1 injections during year 3. Among the 9 (23%) TREX subjects transitioned to PRN, the need for ongoing anti-vascular endothelial growth factor retreatments was small, with 4 (4%) intravitreal injections being delivered among 106 PRN visits; this subgroup displayed an inferior BCVA trajectory compared with the remainder of subjects. Outcomes among subjects continued on TREX were more favourable, with a mean gain of +5.0 letters at M36. CONCLUSIONS: Upon transition to PRN, subjects randomised to monthly dosing experienced a decline in BCVA. Among subjects initially randomised to TREX who transitioned to PRN after achieving a 12-week interval between visits, the overall need for additional treatment was low. TRIAL REGISTRATION NUMBER: NCT01748292, Results.

Relationship Between Visual Acuity and Retinal Thickness During Anti-Vascular Endothelial Growth Factor Therapy for Retinal Diseases.

PURPOSE: To investigate the relationship between best-corrected visual acuity (BCVA) and central retinal thickness (CRT) in eyes receiving ranibizumab for 3 common retinal diseases. DESIGN: Retrospective analysis of clinical trial data. METHODS: Early Treatment Diabetic Retinopathy Study BCVA and spectral-domain optical coherence tomography-measured CRT of 387 eyes of 345 patients enrolled in 6 prospective clinical trials for management of neovascular age-related macular degeneration (AMD), diabetic macular edema (DME), and retinal vein occlusion (RVO) were evaluated by Pearson correlation and linear regression. RESULTS: At baseline, there was a small correlation between BCVA and CRT in pooled AMD trial data (r = -0.24). A medium correlation was identified in pooled DME trial data (r = -0.42). No correlation was found in pooled RVO trial data. At month 12, no correlation was found between changes from baseline in BCVA and CRT in pooled AMD trial data. Medium correlations were identified in both pooled DME (r = -0.45) and pooled RVO (r = -0.35) trial data at month 12. Changes in BCVA and CRT associated with edema recurrence upon transition from monthly to pro re nata (PRN) dosing were correlated in AMD (r = -0.27) and RVO (r = -0.72) trials, but not in DME trial data. CONCLUSION: DME demonstrated a convincing relationship between BCVA and CRT. Correlations appear to be more complex in AMD and RVO. At the inflection point between monthly and PRN dosing, when recurrence of edema is anticipated in many patients, CRT appears strongly correlated with loss of BCVA in RVO.

Randomized Trial of Treat-and-Extend versus Monthly Dosing for Neovascular Age-Related Macular Degeneration: 2-Year Results of the TREX-AMD Study.

PURPOSE: To evaluate a prospective treat-and-extend (TREX) management strategy compared with monthly dosing with intravitreal ranibizumab (Lucentis) in neovascular age-related macular degeneration (AMD). DESIGN: Prospective, randomized, multicenter clinical trial. PARTICIPANTS: Sixty patients with treatment-naïve neovascular AMD randomized 1:2 to monthly or TREX cohorts. METHODS: Patients with Early Treatment Diabetic Retinopathy Study (ETDRS) best-corrected visual acuity (BCVA) of 20/32 to 20/500 (Snellen equivalent) were randomized to receive intravitreal 0.5 mg ranibizumab monthly or, according to a TREX protocol, no less frequently than every 12 weeks. After interval extension, if recurrent exudative disease was identified, this maximum interval between treatments was rechallenged according to a strict prospective protocol. MAIN OUTCOME MEASURE: Change in ETDRS BCVA from baseline. RESULTS: Sixty patients were enrolled and 50 completed month 24, at which point mean ETDRS BCVA letter gains were similar: 10.5 and 8.7 for the monthly and TREX cohorts, respectively (P = 0.64). At month 24, 4 patients (20%) and 12 patients (30%) in the monthly and TREX cohorts, respectively, gained at least 15 letters (P = 0.41). No monthly cohort patient lost more than 2 letters, whereas 5 TREX cohort patients (13%) lost at least 15 letters. Anatomic improvements were similar between the cohorts. Through month 24, the mean number of injections administered was 25.5 (range, 22-27) and 18.6 (range, 10-25) for the monthly and TREX cohorts, respectively (P < 0.001). Among TREX patients completing month 24, 14 (47%) were at an extension interval of 8 weeks or more, and the mean maximum tolerated extension was 8.5 weeks over the course of 2 years. Of the 26 TREX patients (65%) who demonstrated recurrent exudation upon interval extension, the first maximum extension interval was consistent in most eyes (n = 19 [73%]). CONCLUSIONS: The TREX neovascular AMD management protocol used with ranibizumab in the Treat-and-Extend Protocol in Patients with Wet Age-Related Macular Degeneration (TREX-AMD) study resulted in visual and anatomic gains comparable with those obtained with monthly dosing, and most patients randomized to TREX therapy demonstrated a relatively consistent maximum extension interval.

Outcomes With As-Needed Aflibercept and Macular Laser Following the Phase III VISTA DME Trial: ENDURANCE 12-Month Extension Study.

PURPOSE: To determine whether the efficacy and safety achieved with 2.0 mg intravitreal aflibercept injections (IAIs) for diabetic macular edema (DME) during the phase III VISTA DME trial were maintained with individualized, as-needed treatment. DESIGN: Phase IV, multicenter, open-label extension study. METHODS: Sixty patients completing VISTA DME elected to enter the ENDURANCE extension study. All patients received IAIs in the presence of clinically relevant DME. Patients were observed at 4-, 8-, or 12-week intervals depending on the need for treatment. Main outcome measures were mean IAIs given through month 12 (M12), the proportion of patients receiving no IAIs, and the role of macular laser in decreasing treatment burden among patients requiring ongoing IAIs. RESULTS: A mean of 4.5 IAIs were administered through M12. Eighteen (30%) patients required no IAIs, and among those who met IAI retreatment criteria, a mean of 6.0 IAIs were administered through M12. Best-corrected visual acuity gains achieved during VISTA DME were maintained and stable with individualized dosing during ENDURANCE, fluctuating by <1.5 mean letters from the baseline at all time points. Likewise, mean central retinal thickness remained relatively stable during ENDURANCE. Thirty-seven (62%) patients met macular laser criteria at a mean of 19.5 weeks with no significant difference in the frequency of IAIs before or after macular laser. CONCLUSION: Vision gains achieved during the 3-year VISTA DME trial were maintained through M12 of the ENDURANCE extension study with a reduced treatment frequency, with 30% of patients receiving no IAIs. No significant reduction in IAI frequency was observed after macular laser application.