Multimodal, multitask, multiattention (M3) deep learning detection of reticular pseudodrusen: Toward automated and accessible classification of age-related macular degeneration.

OBJECTIVE: Reticular pseudodrusen (RPD), a key feature of age-related macular degeneration (AMD), are poorly detected by human experts on standard color fundus photography (CFP) and typically require advanced imaging modalities such as fundus autofluorescence (FAF). The objective was to develop and evaluate the performance of a novel multimodal, multitask, multiattention (M3) deep learning framework on RPD detection. MATERIALS AND METHODS: A deep learning framework (M3) was developed to detect RPD presence accurately using CFP alone, FAF alone, or both, employing >8000 CFP-FAF image pairs obtained prospectively (Age-Related Eye Disease Study 2). The M3 framework includes multimodal (detection from single or multiple image modalities), multitask (training different tasks simultaneously to improve generalizability), and multiattention (improving ensembled feature representation) operation. Performance on RPD detection was compared with state-of-the-art deep learning models and 13 ophthalmologists; performance on detection of 2 other AMD features (geographic atrophy and pigmentary abnormalities) was also evaluated. RESULTS: For RPD detection, M3 achieved an area under the receiver-operating characteristic curve (AUROC) of 0.832, 0.931, and 0.933 for CFP alone, FAF alone, and both, respectively. M3 performance on CFP was very substantially superior to human retinal specialists (median F1 score = 0.644 vs 0.350). External validation (the Rotterdam Study) demonstrated high accuracy on CFP alone (AUROC, 0.965). The M3 framework also accurately detected geographic atrophy and pigmentary abnormalities (AUROC, 0.909 and 0.912, respectively), demonstrating its generalizability. CONCLUSIONS: This study demonstrates the successful development, robust evaluation, and external validation of a novel deep learning framework that enables accessible, accurate, and automated AMD diagnosis and prognosis.

[Artificial intelligence for eye care]. / Artificiële intelligentie in de oogzorg.

Technological developments in ophthalmic imaging and artificial intelligence (AI) create new possibilities for diagnostics in eye care. AI has already been applied in ophthalmic diabetes care. AI-systems currently detect diabetic retinopathy in general practice with a high sensitivity and specificity. AI-systems for the screening, monitoring and treatment of age-related macular degeneration and glaucoma are promising and are still being developed. AI-algorithms, however, only perform tasks for which they have been specifically trained and highly depend on the data and reference-standard that were used to train the system in identifying a certain abnormality or disease. How the data and the gold standard were established and determined, influences the performance of the algorithm. Furthermore, interpretability of deep learning algorithms is still an ongoing issue. By highlighting on images the areas that were critical for the decision of the algorithm, users can gain more insight into how algorithms come to a particular result.

Performance of Classification Systems for Age-Related Macular Degeneration in the Rotterdam Study.

Purpose: To compare frequently used classification systems for age-related macular degeneration (AMD) in their abilty to predict late AMD. Methods: In total, 9066 participants from the population-based Rotterdam Study were followed up for progression of AMD during a study period up to 30 years. AMD lesions were graded on color fundus photographs after confirmation on other image modalities and grouped at baseline according to six classification systems. Late AMD was defined as geographic atrophy or choroidal neovascularization. Incidence rate (IR) and cumulative incidence (CuI) of late AMD were calculated, and Kaplan-Meier plots and area under the operating characteristics curves (AUCs) were constructed. Results: A total of 186 persons developed incident late AMD during a mean follow-up time of 8.7 years. The AREDS simplified scale showed the highest IR for late AMD at 104 cases/1000 py for ages <75 years. The Rotterdam classification showed the highest IR at 89 cases/1000 py >75 years. The 3-Continent harmonization classification provided the most stable progression. Drusen area >10% ETDRS grid (hazard ratio 30.05, 95% confidence interval [CI] 19.25-46.91) was most prognostic of progression. The highest AUC of late AMD (0.8372, 95% CI: 0.8070-0.8673) was achieved when all AMD features present at baseline were included. Conclusions: Highest turnover rates from intermediate to late AMD were provided by the AREDS simplified scale and the Rotterdam classification. The 3-Continent harmonization classification showed the most stable progression. All features, especially drusen area, contribute to late AMD prediction. Translational Relevance: Findings will help stakeholders select appropriate classification systems for screening, deep learning algorithms, or trials.

Cytokine and chemokine dynamics differ between rats and mice after collagen implantation.

Implanted scaffold materials induce an inflammatory reaction known as the 'foreign body reaction' (FBR). We hypothesized that the observed difference in FBR between rats and mice correlate with different expression dynamics of cytokines and chemokines, which are key orchestrators of the FBR. After implantation of hexamethylene diisocyanate cross-linked dermal sheep collagen, the overall gene expression pattern of IL-1, IL-6, IL-10, TNFalpha, CXCL1/KC, CXCL2/MIP2 and CCL2/MCP1 was roughly similar for the two species. During the onset of the FBR these genes were maximally expressed in rats and mice, after which the expression decreased to basal levels. The expression of CCL3/MIP1 alpha had a similar course, yet it increased after the progression phase of the FBR in both species. The expression of cytokines and chemokines in sham-operated animals was low throughout, showing that the implanted material by itself exerted the changes in gene expression of the invading cells. During the progression, genes encoding the PMN attractants CXCL1/KC and CXCL2/MIP2 were more highly expressed in mice than in rats, which would explain the prolonged presence of PMNs in mice during the FBR. Additionally, the strong induction of IFN gamma in rats coincided with a higher phagocytotic activity by macrophages. Throughout the FBR, the expression of TGFbeta was constitutive and high in both species, but increased in mice during the progression phase. This could explain the extensive stroma formation during the murine FBR. Unexpectedly, the stronger expression of TNFalpha and CCL3/MIP1 alpha in mice, did not result in high macrophage attraction or phagocytosis of the implanted collagen disks.

Cellular and molecular dynamics in the foreign body reaction.

Intracorporally implanted materials, such as medical devices, will provoke the body to initiate an inflammatory reaction. This inflammatory reaction to implanted materials is known as the foreign body reaction (FBR) and is characterized by 3 distinct phases: onset, progression, and resolution. The FBR proceeds in the creation of a dynamic microenvironment that is spatially well organized. The progression of the FBR is regulated by soluble mediators, such as cytokines, chemokines, and matrix metalloproteinases (MMPs), which are produced locally by tissue cells and infiltrated inflammatory cells. These soluble mediators orchestrate the cascade of cellular processes in the microenvironment that accompanies the FBR, consisting of cellular activation, angiogenesis, extravasation, migration, phagocytosis, and, finally, fibrosis. The nature of the FBR requires that the soluble mediators act in a spatial and temporally regulated manner as well. This regulation is well known for several inflammatory processes, but scarce knowledge exists about the intricate relationship between the FBR and the expression of soluble mediators. This review discusses the key processes during the initiation, progression, and resolution phase, with emphasis on the role of soluble mediators. Besides other sites of implantation, we focus on the subcutaneous implantation model.

The correlation between difference in foreign body reaction between implant locations and cytokine and MMP expression.

The foreign body reaction (FBR) differs between subcutaneously and supra-epicardially implanted materials. We hypothesize that this is a result of differences in cytokine, chemokine and matrix metalloproteinase (MMP) dynamics. Therefore we applied collagen disks subcutaneously and on the epicardium in mice and analyzed the FBR from day 1 to 21. Both the influx of leukocytes and implant degradation were higher in supra-epicardially implanted collagen than in subcutaneously implanted material. This correlated with a higher gene expression of pro-inflammatory cytokines such as IL-1 and IL-6, and a lower expression of the anti-inflammatory cytokine IL-10. Furthermore, the higher supra-epicardial expression of PMN attractants CXCL1/KC and CXCL2/MIP2 correlated with a higher and prolonged PMN influx. The gene expression levels of collagen degrading MMPs, i.e. MMP8, MMP13 and MMP14 were similar in subcutaneous and supra-epicardial disks. However, the activity of these enzymes was markedly higher supra-epicardially. In addition, the MMP9 expression was higher supra-epicardially, suggesting a role for this enzyme in the degradation process. In conclusion, a strong pro-inflammatory milieu is generated after supra-epicardial implantation that enables prolonged PMN presence and activation. This, together with the high supra-epicardial MMP9 level, could explain the observed difference in Col-I degradation between locations.

The modulation of angiogenesis in the foreign body response by the poxviral protein M-T7.

The foreign body response is characterized by enhanced recruitment of inflammatory cells. As the directional movement of cells is controlled by chemokines, disruption of the chemokine network would be an attractive approach to improve biocompatibility of an implanted material. The sequestration of chemokines by cell surface-expressed glycosaminoglycans (GAGs) is vital for in vivo chemokine activity. The myxoma virus encodes a soluble protein, M-T7, that interacts with conserved GAG-binding domains of chemokines to block chemokine-mediated leukocyte recruitment. We hypothesized that M-T7 might also affect the function of other inflammation-associated proteins in addition to chemokines that bind to GAG. In our studies, we focussed on the modulation of the GAG-binding molecules macrophage chemoattractant protein-1 (MCP-1) and vascular endothelial growth factor-164 (VEGF164) in the inflammatory reaction against subcutaneously implanted degradable cross-linked dermal sheep collagen discs in AO rats. Genetic delivery of M-T7 delays the influx of macrophages into the collagen discs. In addition, angiogenesis around the implanted material was reduced. The discs revealed reduced levels of rat MCP-1 and rat VEGF164. This was not due to down regulation of transcription of the genes that encode MCP-1 and VEGF164. Our in vivo observations suggest that, in addition to chemokines such as MCP-1, M-T7 neutralizes VEGF164.