astro-ph.GAApr 8, 2026
SaveEuclid Quick Data Release (Q1). AgileLens: A scalable CNN-based pipeline for strong gravitational lens identification
Organizations: Department of Physics \& Astronomy, University of California Irvine, Irvine CA 92697, USA · University of Southern California, 3551 Trousdale Parkway, Los Angeles, CA 90089, USA · Dipartimento di Fisica "Aldo Pontremoli", Università degli Studi di Milano, Via Celoria 16, 20133 Milano, Italy · INAF-IASF Milano, Via Alfonso Corti 12, 20133 Milano, Italy ·
Abstract
We present an end-to-end, iterative pipeline for efficient identification of strong galaxy--galaxy lensing systems, applied to the Euclid Q1 imaging data. Starting from VIS catalogues, we reject point sources, apply a magnitude cut (I 24) on deflectors, and run a pixel-level artefact/noise filter to build 96 96 pix cutouts; VIS+NISP colour composites are constructed with a VIS-anchored luminance scheme that preserves VIS morphology and NISP colour contrast. A VIS-only seed classifier supplies clear positives and typical impostors, from which we curate a morphology-balanced negative set and augment scarce positives. Among the six CNNs studied initially, a modified VGG16 (GlobalAveragePooling + 256/128 dense layers with the last nine layers trainable) performs best; the training set grows from 27 seed lenses (augmented to 1809) plus 2000 negatives to a colour dataset of 30,686 images. After three rounds of iterative fine-tuning, human grading of the top 4000 candidates ranked by the final model yields 441 Grade A/B candidate lensing systems, including 311 overlapping with the existing Q1 strong-lens catalogue, and 130 additional A/B candidates (9 As and 121 Bs) not previously reported. Independently, the model recovers 740 out of 905 (81.8%) candidate Q1 lenses within its top 20,000 predictions, considering off-centred samples. Candidates span I 17--24 AB mag (median 21.3 AB mag) and are redder in Y--H than the parent population, consistent with massive early-type deflectors. Each training iteration required a week for a small team, and the approach easily scales to future Euclid releases; future work will calibrate the selection function via lens injection, extend recall through uncertainty-aware active learning, explore multi-scale or attention-based neural networks with fast post-hoc vetters that incorporate lens models into the classification.
Figures & tables
Figure 1: Schematic representation of the iterative training pipeline. Each iteration comprises three steps: Step 1: inspection of the top-ranked candidate objects in the model output; Step 2: update of the training dataset through the inclusion of newly identified gravitational lens candidates and false positives; and Step 3: fine-tuning of the model with the updated dataset. The process is repeated until convergence, i.e., when no additional lenses are identified among the top-ranked candidates.
Figure 2: Examples of randomly selected artefacts (displayed in -band images) not removed by the SPURIOUS_PROB filter. In almost all cases, these are bright stars with centrally saturated pixels that are captured by L2 image flags.
Figure 3: Examples of VIS band images images containing defects made of large and small pixel areas, and noisy images that are removed by our artefact removal algorithm.
| Point. | Point.+Spur. | Point.+Artefact | |
| 0–23 | 2 594 180 | 2 361 825 | 2 170 933 |
| 23–24 | 2 867 796 | 2 025 908 | 1 265 417 |
Table 1: Number of detected sources per magnitude bin after successive quality filters.
Figure 4: Examples of confident strong lens images used for building the initial Euclid positive training set based out of VIS-only detection scheme used at Step 1. We augmented these initial detections from 27 candidate lenses to a total of true positive cases using a combination of rotation, translation, and colour transformations.
Figure 5: Examples of false positives identified from the output of pre-trained model. These false positives consist of four typical types: elliptical galaxies (top row), spiral galaxies (second row), edge-on spiral galaxies (third row), and images containing diffraction spikes of a bright star close to a galaxy or another source (buttom row).
| Initial set | Component | Size |
| Positive | 27 lens candidates, augmented 67-fold | 1809 |
| Negative | Balanced sample of: spiral-arm galaxies elliptical galaxies edge-on galaxies diffraction spikes | 2000 |
Table 2: Components and sizes of the initial Euclid imaging training set later used to fine-tune models and evaluate their performance.
| Model name | Lenses found | Lens fraction |
| VGG16 ( Simonyan and Zisserman, 2015 ) | 34 | 0.27 |
| Xception ( Chollet, 2017 ) | 27 | 0.21 |
| ResNet50 ( He et al., 2015 ) | 24 | 0.19 |
| InceptionV3 ( Szegedy et al., 2015 ) | 23 | 0.18 |
| MobileNetV2 ( Howard et al., 2017 ) | 8 | 0.06 |
| EfficientNetB0 ( Tan and Le, 2020 ) | 6 | 0.05 |
Table 3: Performance of candidate CNN architectures fine-tuned on the initial training set.
Figure 6: Architecture of the modified VGG16 convolutional neural network used for lens classification. The original fully connected layers were replaced with a customised final structure, labelled in purple. This includes a GlobalAveragePooling2D layer to reduce spatial dimensions, followed by two dense layers with 256 and 128 units to capture high-level features, and a final sigmoid layer for binary classification.
Figure 7: Three newly identified (relative to the previous iterative round) Grade A strong lens candidates from the fine-tuned model in the second round that exhibit atypical "lemon” and white colouring compared to most other lensing systems that show a colour difference between the foreground lens and the background source. The first system (left) is NGC 6505, a complete Einstein ring, discussed by O’Riordan et al. (2025) based on the visual identification of the ring.
Figure 8: Three rounds of the iterative model retraining process cross-validated with 905 sources (containing off-centred samples) from this work and SLDE A/B/F. The left-hand plot denotes the first iteration trained on the initial training dataset selected by the pre-trained model. The blue curve shows recall, that is, where in the ranked predictions we could find each of the confirmed lens candidates. The green curve denotes the recall of a perfect classifier, while the orange line denotes a completely random classifier. The red curve denotes the precision of the current model, that is, how well the model moves non-lenses to lower rankings. Throughout development, we found that the actual recall curve mostly overlaps with the perfect classifier in the top 100 predicted candidates.
| Processed classification | Vote No. | Fraction (%) |
| Grade X: not a lens | 3095 | 77.4 |
| Grade C: lens-like feature | 464 | 11.6 |
| Grade B: probable lens | 314 | 7.9 |
| Grade A: confidently a lens | 111 | 2.8 |
| Grade A+: lens of science value | 16 | 0.4 |
| Total | 4000 | 100.0 |
Table 4: Distribution of post-processed classification outcomes from Euclid expert ratings for the model’s top 4000 predicted objects.
Figure 9: Light blue bars shows the "raw score” of the top 4000 cutouts ranked by the model, and dark blue bars represent the identified A/B lens candidates within the 4000 cutouts. Orange line shows the fraction of lens candidates, computed as the ratio of dark blue over light blue in each bin. On the lower "raw score” end, the fraction of lenses declines to less than 0.05. This justifies the threshold for visual inspections as lenses beyond top 4000 are too rare for effective human inspections.
Figure 10: Pie chart showing distribution of 130 new lenses identified and lenses previously discovered by SLDE A in the top 4000 predictions. Each grade is a concatenation of grade in this work and in SLDE A. For instance, the BA slice indicates that 75 (13.1%) lenses received grade B in this work but grade A in SLDE A.
Figure 11: Four new lenses within the flux magnitude thresholds used in SLDE A but not identified as lens candidates in that work. Objects outlined in orange were identified as lenses by previous models in SLDE C but rejected by citizen scientists in Space Warps, a classification project hosted on Zooniverse platform. The other two were poorly ranked by SLDE C machine learning models.
Figure 12: Distribution of model prediction "raw scores” for candidates grouped by human-assigned grade: “non-lens”, C, B, and A. Each subplot shows a histogram of prediction values, with red dashed lines indicating the mean and yellow dotted lines indicating the median for that grade. As the grade increases, the mean and median of the prediction values shift towards the right, indicating a higher prediction value. Thus, we can infer that the prediction value from our model correlates with a higher grade.
Figure 13: 20 false positives within the top 4000 classifications of the model. Within these we can see an example with a diffraction spike, several elliptical and spiral galaxies, with many containing evidence of recent mergers or an ongoing merger in the form of a tidal stream, as well as cut-outs containing multiple objects.
Figure 14: Histogram of object counts as a function of detection magnitude. Top: Blue bars represent the full set of inspected candidates with ; Middle: dark and light colour corresponds to the grade A and grade B lens candidates respectively, and the background bars are the sum of the two. For reference, we also show the forecast strong lens magnitude distribution from Collett (2015) normalised to Euclid Q1 area. Bottom: The ratio of lens candidates inside the parent population as a function of magnitude.
Figure 15: colour vs. magnitude distribution for model-discovered lenses and the parent population. Orange triangles represent the lenses previously found in SLDE A, while green circles represent lenses not found in SLDE A. The blue contours represent the distribution of the parent population, with contour lines at , , , and confidence levels.
Figure 16: "Raw score” distribution of all cut-outs and the identified lenses. The dashed line represents the inspection threshold of the top 4000 "raw scores”; all cutouts to the left of the dashed line were not inspected.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 17: Centred lens candidates that received lower grades in this work compared with SLDE A. Panel (a) displays objects that received Grade A in SLDE A but were graded as Grade B in this work. Panel (b) displays objects that received Grade B in SLDE A but were graded as Grade C in this work. Panel (c) displays objects that received Grade A in SLDE A but were graded as Grade C in this work.
Figure 18: Centred lens candidates that received higher grades in this work compared with SLDE A. Panel (a) displays objects that received Grade A in this work but were graded as Grade B in SLDE A. Panel (b) displays objects that received Grade B in this work but were graded as C in SLDE A. Panel (c) displays objects that received Grade A in this work but were graded as C in SLDE A.
Figure 19: A panel of the 100 experimental hand-drawn lens simulations used to enhance the diversity of the positive training dataset in the second round of the iterative training process. These images were created using Procreate and were based on visual features commonly exhibited among found lenses. These features include the red–orange colour of the lensing galaxy, purple colour of the background galaxy, various observed lensing configurations, etc.
Figure 20: (1–80) 367 classified grade A/B lens candidates. New lens candidates not included in Euclid Collaboration: Walmsley et al. (2025) are outlined in green. Images displayed are slightly enhanced with an arcsinh function; however, this function was only applied for display – the model itself was applied to linearly stretched images in training and inference.
Figure 21: (81–160) 367 classified grade A/B lens candidates. New lens candidates not included in SLDE A are outlined in green. Images displayed are slightly enhanced with an arcsinh function; however, this function was only applied for display – the model itself is applied to linearly stretched images in training and inference.
Figure 22: (161–240) 367 classified grade A/B lens candidates. New lens candidates not included in SLDE A are outlined in green. Images displayed are slightly enhanced with an arcsinh function; however, this function was only applied for display – the model itself was applied to linearly stretched images in training and inference.
Figure 23: (241–320) 367 classified grade A/B lens candidates. New lens candidates not included in SLDE A are outlined in green. Images displayed are slightly enhanced with an arcsinh function; however, this function was only applied for display – the model itself was applied to linearly stretched images in training and inference.
Figure 24: (321–367) 367 classified grade A/B lens candidates. New lens candidates not included in SLDE A are outlined in green. Images displayed are slightly enhanced with an arcsinh function; however, this function was only applied for display – the model itself was applied to linearly stretched images in training and inference.
| OBJ NAME | RA | Dec | GRADE | RAW SCORE | NOTES | ||||
| (1) | (2) | (3) | (4) | (5) | (6) | (7) | (8) | (9) | (10) |
| \csvreader [ head to column names, late after line= | |||||||||
| , separator=semicolon, filter=<30 ]tables/master_table_v7.csv \objname ° \dec \VIS \Y ˝ \grade \rawscore \notes |
Table 5: A few example entries of the master table of classified lenses depicted in D . Currently depicting the top 18 lenses from Fig. 20 . The properties displayed are (1): the object id/name for each candidate (2): right ascension (3): declination (4): band magnitude, using FLUX_DETECTION_TOTAL (5) band magnitude, found using FLUX_Y_2FWHM_APER (6): band magnitude, found using FLUX_J_2FWHM_APER (7): band magnitude, found using FLUX_H_2FWHM_APER (8) The processed grades of human experts classification. If there are two letters, then the object was previously identified and classified, with grade represented by the second letter (9): The "raw score” output from the model in this work (10): citations and any additional notes. Two undisplayed properties are included in the full electronically available table in additional resources: whether the lens is new; and an alternative name for each object composed of the mosaic tile and cutout id. Under notes, L25: Euclid Collaboration: Lines et al. (2025b) , W25: Euclid Collaboration: Walmsley et al. (2025) , R25: Euclid Collaboration: Rojas et al. (2025) , and E25: Euclid Collaboration: Ecker et al. (2025) . R25 and E25 not shown here.
More in this topic
Convolutional Neural Networks
Convolutional neural networks (CNNs) learn spatial and temporal patterns using shared filters, with applications from image recognition to scientific prediction.
See all 401 papers