Overlap-Aware Segmentation (OASIS)

• The paper introduces OASIS, a segmentation-regression framework that adjusts the loss function to emphasize overlapping regions and improve faint ER signal reconstruction in low-energy imaging.
• The method employs a weighted loss formulation with specific channel and region weights, yielding improved median percent error and IoU compared to unweighted baselines.
• The framework is generalizable to various imaging domains such as astronomy, microscopy, and medical imaging, offering an overlap-aware approach without altering the underlying model architecture.

Searching arXiv for OASIS and related overlap-aware segmentation papers. Overlap-Aware Segmentation of ImageS (OASIS) denotes a segmentation-regression framework in which the training loss explicitly prioritizes regions of object overlap, rather than treating all pixels as equally informative. In the formulation introduced for scientific imaging, the network outputs object-specific intensity maps and is optimized to recover obscured signals in overlap-dominated scenes, with the MIGDAL experiment serving as the primary demonstration case (Schueler et al., 7 Oct 2025). Within a broader overlap-aware segmentation lineage, earlier geometry-driven work addressed heavily overlapping convex objects by decomposing irregular contours, regrouping contour fragments, and assigning shape-specific geometric entities, particularly in crystal microscopy with mixed rod/needle-shaped and circular/spherical/prismatic objects (Abhinav et al., 2017).

1. Conceptual scope and problem setting

OASIS is designed for scientific images in which pixels represent physical intensities rather than only semantic labels. The target is therefore not merely pixel classification, but decomposition of an input image into object-specific intensity maps. This becomes difficult when one object is faint and partially or mostly hidden by another, because ordinary segmentation-regression typically optimizes all pixels equally, even though attribution is most ambiguous in overlap regions (Schueler et al., 7 Oct 2025).

The MIGDAL application exemplifies the intended regime. The reconstruction target is a very faint electron recoil (ER) track that is often heavily buried within an orders-of-magnitude brighter nuclear recoil (NR) track. In the low-energy search region, approximately $4$–$15$ keV ER energy, the ER track can be almost completely buried. Because the Migdal cross section rises as ER energy decreases, the hardest reconstructions are also the most scientifically important (Schueler et al., 7 Oct 2025).

A central clarification is that OASIS is a framework, not just a network. It retains a standard segmentation-regression backbone, but alters the loss so that specific channels and specific spatial regions can be emphasized. In the MIGDAL case, the network predicts $k=2$ intensity maps, one for ER and one for NR. The distinctive element is therefore the weighting strategy, not a replacement of the underlying model class (Schueler et al., 7 Oct 2025).

2. Weighted loss formulation

The OASIS objective combines a weighted segmentation-regression term with a smoothness penalty:

$$\mathcal{L} = \alpha_{1}\,\mathcal{L}_{\text{seg-reg}} + \alpha_{2}\,\mathcal{L}_{\mathrm{smooth}},$$

with $\alpha_1 = 1$ and $\alpha_2 = 0.01$ (Schueler et al., 7 Oct 2025).

The main term is a weighted mean absolute error over channels and pixels:

$$\mathcal{L}_{\text{seg-reg}} = \frac{\sum_{c=1}^{k} w_c \sum_{(x,y)\in B} W_{\mathrm{region}(x,y)}\,\left|\hat{\mathcal{I}}_c(x,y)-\mathcal{I}_c(x,y)\right|} {\sum_{c=1}^{k} w_c \sum_{(x,y)\in B} W_{\mathrm{region}(x,y)}}.$$

Here, $c$ indexes the object class or channel, $\hat{\mathcal{I}}_c(x,y)$ is the predicted intensity, $\mathcal{I}_c(x,y)$ is the ground truth, $15$0 is an optional bounding box region of interest, $15$1 is the channel weight, and $15$2 is the region weight (Schueler et al., 7 Oct 2025).

The smoothness regularizer is total-variation-like:

$15$3

Its role is to discourage unnaturally jagged reconstructions (Schueler et al., 7 Oct 2025).

The overlap weighting is derived from label structure. In the general discussion, overlap regions are those where multiple object channels have nonzero pixel intersection; with three objects, for example, distinct weights could be assigned to $15$4, $15$5, $15$6, and $15$7. In the MIGDAL implementation, the region weights are $15$8 for ER-only pixels, $15$9 for NR-only pixels, and $k=2$0 for overlap pixels where both ER and NR have nonzero truth intensity. Channel weights are $k=2$1 and $k=2$2. The unweighted baseline sets all channel and region weights to $k=2$3 (Schueler et al., 7 Oct 2025).

This weighting scheme encodes the claim that overlap pixels are both the most ambiguous and the most informative. A plausible implication is that OASIS operationalizes overlap-awareness as a supervision policy: ambiguity is not removed from the data, but is upweighted during optimization.

3. Backbone, data, and optimization

The backbone used in the MIGDAL study is a U-Net with four encoder stages, a bottleneck, a mirrored decoder, skip connections, group normalization, SiLU activations, bilinear upsampling in the decoder, and a final $k=2$4 convolution plus Softplus output activation. The architecture uses input size $k=2$5, base channel depth $k=2$6, and output channels $k=2$7 (Schueler et al., 7 Oct 2025).

Training uses $k=2$8 hybrid signal images containing one ER plus one real NR, and $k=2$9 NR-only images. The NR-only images are included to prevent the model from hallucinating an ER when none is present. Preprocessing follows the MIGDAL search pipeline: $$\mathcal{L} = \alpha_{1}\,\mathcal{L}_{\text{seg-reg}} + \alpha_{2}\,\mathcal{L}_{\mathrm{smooth}},$$0 binning, Gaussian smoothing with a $$\mathcal{L} = \alpha_{1}\,\mathcal{L}_{\text{seg-reg}} + \alpha_{2}\,\mathcal{L}_{\mathrm{smooth}},$$1 kernel and $$\mathcal{L} = \alpha_{1}\,\mathcal{L}_{\text{seg-reg}} + \alpha_{2}\,\mathcal{L}_{\mathrm{smooth}},$$2, and negative pixels set to $$\mathcal{L} = \alpha_{1}\,\mathcal{L}_{\text{seg-reg}} + \alpha_{2}\,\mathcal{L}_{\mathrm{smooth}},$$3. The input is then log-scaled as

$$\mathcal{L} = \alpha_{1}\,\mathcal{L}_{\text{seg-reg}} + \alpha_{2}\,\mathcal{L}_{\mathrm{smooth}},$$4

where $$\mathcal{L} = \alpha_{1}\,\mathcal{L}_{\text{seg-reg}} + \alpha_{2}\,\mathcal{L}_{\mathrm{smooth}},$$5 is the $$\mathcal{L} = \alpha_{1}\,\mathcal{L}_{\text{seg-reg}} + \alpha_{2}\,\mathcal{L}_{\mathrm{smooth}},$$6th percentile of the pixel intensities. This compresses dynamic range while preserving faint ER structure (Schueler et al., 7 Oct 2025).

Optimization uses a $$\mathcal{L} = \alpha_{1}\,\mathcal{L}_{\text{seg-reg}} + \alpha_{2}\,\mathcal{L}_{\mathrm{smooth}},$$7 train/validation split, batch size $$\mathcal{L} = \alpha_{1}\,\mathcal{L}_{\text{seg-reg}} + \alpha_{2}\,\mathcal{L}_{\mathrm{smooth}},$$8, AdamW, learning rate $$\mathcal{L} = \alpha_{1}\,\mathcal{L}_{\text{seg-reg}} + \alpha_{2}\,\mathcal{L}_{\mathrm{smooth}},$$9, weight decay $\alpha_1 = 1$0, and early stopping patience of $\alpha_1 = 1$1 epochs. The implementation is in PyTorch and is accelerated on an Nvidia RTX 3090 (Schueler et al., 7 Oct 2025).

The experimental apparatus in the MIGDAL application consists of fast neutrons from a D-D generator, a low-pressure optical TPC, $\alpha_1 = 1$2 Torr CF$\alpha_1 = 1$3, and a Hamamatsu ORCA Quest qCMOS camera. This specific setting matters because the overlap problem is not only geometric but also radiometric: a faint ER track must be separated from a much brighter NR track in a detector image whose pixel values correspond to physical quantities (Schueler et al., 7 Oct 2025).

4. Evaluation protocol and reported performance

The principal scalar for ER reconstruction is the track intensity

$\alpha_1 = 1$4

with percent error

$\alpha_1 = 1$5

Topological agreement is measured with ER intersection-over-union:

$\alpha_1 = 1$6

Angular consistency is assessed by fitting principal curves to predicted and truth ER maps and comparing principal axes, yielding an angle difference $\alpha_1 = 1$7. For NR-only events, a false positive ER is defined by reconstructed ER visible energy exceeding $\alpha_1 = 1$8 keV (Schueler et al., 7 Oct 2025).

The key quantitative comparison is weighted versus unweighted training in the $\alpha_1 = 1$9–$\alpha_2 = 0.01$0 keV ER bin, where overlap is most severe. Median percent error improves from $\alpha_2 = 0.01$1 to $\alpha_2 = 0.01$2, and median IoU improves from $\alpha_2 = 0.01$3 to $\alpha_2 = 0.01$4. The false positive rate on NR-only test events, using the $\alpha_2 = 0.01$5 keV threshold, changes from $\alpha_2 = 0.01$6 for the unweighted model to $\alpha_2 = 0.01$7 for the weighted model (Schueler et al., 7 Oct 2025).

Event-level examples further illustrate the spread of outcomes: one good event with $\alpha_2 = 0.01$8 and $\alpha_2 = 0.01$9; another with $$\mathcal{L}_{\text{seg-reg}} = \frac{\sum_{c=1}^{k} w_c \sum_{(x,y)\in B} W_{\mathrm{region}(x,y)}\,\left|\hat{\mathcal{I}}_c(x,y)-\mathcal{I}_c(x,y)\right|} {\sum_{c=1}^{k} w_c \sum_{(x,y)\in B} W_{\mathrm{region}(x,y)}}.$$0 and $$\mathcal{L}_{\text{seg-reg}} = \frac{\sum_{c=1}^{k} w_c \sum_{(x,y)\in B} W_{\mathrm{region}(x,y)}\,\left|\hat{\mathcal{I}}_c(x,y)-\mathcal{I}_c(x,y)\right|} {\sum_{c=1}^{k} w_c \sum_{(x,y)\in B} W_{\mathrm{region}(x,y)}}.$$1; one difficult event with $$\mathcal{L}_{\text{seg-reg}} = \frac{\sum_{c=1}^{k} w_c \sum_{(x,y)\in B} W_{\mathrm{region}(x,y)}\,\left|\hat{\mathcal{I}}_c(x,y)-\mathcal{I}_c(x,y)\right|} {\sum_{c=1}^{k} w_c \sum_{(x,y)\in B} W_{\mathrm{region}(x,y)}}.$$2 and $$\mathcal{L}_{\text{seg-reg}} = \frac{\sum_{c=1}^{k} w_c \sum_{(x,y)\in B} W_{\mathrm{region}(x,y)}\,\left|\hat{\mathcal{I}}_c(x,y)-\mathcal{I}_c(x,y)\right|} {\sum_{c=1}^{k} w_c \sum_{(x,y)\in B} W_{\mathrm{region}(x,y)}}.$$3; and an NR-only case with $$\mathcal{L}_{\text{seg-reg}} = \frac{\sum_{c=1}^{k} w_c \sum_{(x,y)\in B} W_{\mathrm{region}(x,y)}\,\left|\hat{\mathcal{I}}_c(x,y)-\mathcal{I}_c(x,y)\right|} {\sum_{c=1}^{k} w_c \sum_{(x,y)\in B} W_{\mathrm{region}(x,y)}}.$$4 eV, effectively zero. For the signal test set in the $$\mathcal{L}_{\text{seg-reg}} = \frac{\sum_{c=1}^{k} w_c \sum_{(x,y)\in B} W_{\mathrm{region}(x,y)}\,\left|\hat{\mathcal{I}}_c(x,y)-\mathcal{I}_c(x,y)\right|} {\sum_{c=1}^{k} w_c \sum_{(x,y)\in B} W_{\mathrm{region}(x,y)}}.$$5–$$\mathcal{L}_{\text{seg-reg}} = \frac{\sum_{c=1}^{k} w_c \sum_{(x,y)\in B} W_{\mathrm{region}(x,y)}\,\left|\hat{\mathcal{I}}_c(x,y)-\mathcal{I}_c(x,y)\right|} {\sum_{c=1}^{k} w_c \sum_{(x,y)\in B} W_{\mathrm{region}(x,y)}}.$$6 keV range, $$\mathcal{L}_{\text{seg-reg}} = \frac{\sum_{c=1}^{k} w_c \sum_{(x,y)\in B} W_{\mathrm{region}(x,y)}\,\left|\hat{\mathcal{I}}_c(x,y)-\mathcal{I}_c(x,y)\right|} {\sum_{c=1}^{k} w_c \sum_{(x,y)\in B} W_{\mathrm{region}(x,y)}}.$$7 events are considered and principal curves are successfully fit for $$\mathcal{L}_{\text{seg-reg}} = \frac{\sum_{c=1}^{k} w_c \sum_{(x,y)\in B} W_{\mathrm{region}(x,y)}\,\left|\hat{\mathcal{I}}_c(x,y)-\mathcal{I}_c(x,y)\right|} {\sum_{c=1}^{k} w_c \sum_{(x,y)\in B} W_{\mathrm{region}(x,y)}}.$$8, about $$\mathcal{L}_{\text{seg-reg}} = \frac{\sum_{c=1}^{k} w_c \sum_{(x,y)\in B} W_{\mathrm{region}(x,y)}\,\left|\hat{\mathcal{I}}_c(x,y)-\mathcal{I}_c(x,y)\right|} {\sum_{c=1}^{k} w_c \sum_{(x,y)\in B} W_{\mathrm{region}(x,y)}}.$$9. Median angular consistencies are stated to be within about $c$0 across energies $c$1 keV (Schueler et al., 7 Oct 2025).

The performance pattern supports the paper’s central interpretation: weighting helps most where the ER is shortest, faintest, and most buried. At the same time, the increase in false positive rate shows that improved sensitivity to obscured ERs comes with a modest increase in over-attribution.

5. Limitations, assumptions, and generalization

Several assumptions are explicit. OASIS depends on having a meaningful way to define overlap regions from labels or simulation. The weights are hand-chosen and were selected from preliminary comparisons. The MIGDAL study uses hybrid simulated events composed of real NR tracks and simulated ER tracks, which improves realism but does not constitute a fully end-to-end real-data truth benchmark. The angular reconstruction further uses an idealized principal-curve method and, for ERs, truth vertex information from simulation; in real deployment, vertex finding will depend on a separate NR vertex algorithm (Schueler et al., 7 Oct 2025).

These conditions delimit what OASIS establishes. It demonstrates that one can improve source separation in overlap-dominated scientific imaging without redesigning the network architecture, by reweighting the loss so that channels and regions associated with ambiguity receive more supervision. The practical implication is methodological rather than domain-specific: a standard segmentation-regression model can be made overlap-aware by changing the objective function (Schueler et al., 7 Oct 2025).

The framework is explicitly presented as generalizable to astronomy, such as galaxy deblending, microscopy, medical imaging, particle detectors and TPCs, and scientific imaging more broadly where pixels correspond to intensities and overlap obscures sources. Multi-band inputs and $c$2D voxel data are also described as straightforward extensions. Code is stated to be openly available at https://github.com/jschuel/OASIS, and related open software in the workflow includes migYOLO, PyTorch, YOLOv8, and principal-curve fitting in R (Schueler et al., 7 Oct 2025).

A common misconception would be to treat OASIS as a specialized detector architecture for MIGDAL alone. The paper instead defines it as a weighting framework layered on top of standard segmentation-regression.

6. Relation to earlier overlap-aware segmentation of obscured objects

Prior to the OASIS framework, overlap-aware segmentation was developed in a geometry-driven form for images containing convex objects present in multiple shapes with a high degree of overlap. "Image Segmentation of Multi-Shaped Overlapping Objects" formulates the task as contour extraction, contour mapping, and shape identification, with particular attention to crystal microscopy where rod/needle-shaped and circular/spherical/prismatic crystals overlap so densely that true boundaries must be inferred from contour geometry rather than image gradients (Abhinav et al., 2017).

That method proceeds in two main stages: Contour Region Detection (CRD) and Shape Fitting. Starting from an RGB image, it converts to grayscale, applies median blur, binarizes using Otsu’s method to obtain $c$3, and applies morphological operations to remove noise and fill holes. Visible contours are extracted from the binary image using the border following algorithm of Suzuki et al., represented as

$c$4

Because a single irregular contour may correspond to multiple overlapping objects, the extracted boundary is treated as a contour cluster rather than as a single object outline (Abhinav et al., 2017).

Overlap-awareness is introduced through concave-point detection. Corner points are extracted with the Ramer-Douglas-Peucker (RDP) algorithm,

$c$5

and local orientation is computed from adjacent contour vectors. The sign of the cross product determines $c$6, the net contour orientation is the sign of the summed local orientations, and a corner whose local orientation differs from the net orientation is marked as concave. The resulting set

$c$7

splits the contour cluster into contour segments $c$8, so that

$c$9

This is the explicit mechanism by which overlap-induced indentations become cut points for object separation (Abhinav et al., 2017).

After splitting, the algorithm fits an ellipse $\hat{\mathcal{I}}_c(x,y)$0 to each segment $\hat{\mathcal{I}}_c(x,y)$1 using classical least squares ellipse fitting and extracts major axis, minor axis, orientation angle, and aspect ratio

$\hat{\mathcal{I}}_c(x,y)$2

If $\hat{\mathcal{I}}_c(x,y)$3, the segment is treated as close to a circle and regarded as a unique segment not mapped to others. Otherwise, elongated segments $\hat{\mathcal{I}}_c(x,y)$4 and $\hat{\mathcal{I}}_c(x,y)$5 are grouped when

$\hat{\mathcal{I}}_c(x,y)$6

The grouped set is

$\hat{\mathcal{I}}_c(x,y)$7

with

$\hat{\mathcal{I}}_c(x,y)$8

Each $\hat{\mathcal{I}}_c(x,y)$9 is interpreted as the visible contour of one unique object in the overlap cluster (Abhinav et al., 2017).

Shape assignment is adaptive. In the homogeneous case, ellipse fitting is sufficient. In the heterogeneous case, the method introduces shape classes: rod/needle-shaped and circular/spherical-shaped. Aspect ratio $\mathcal{I}_c(x,y)$0 yields circle-like, aspect ratio $\mathcal{I}_c(x,y)$1 yields rod-like, with the threshold $\mathcal{I}_c(x,y)$2 chosen manually from data analysis. Circular objects are represented by the fitted ellipse, whereas rod-shaped objects are represented by a custom bounding box algorithm because ellipse fitting overestimates length for thin rods. Two rejection filters are then applied: a masking rejection filter that rejects an ellipse if it covers fewer than $\mathcal{I}_c(x,y)$3 black pixels inside its boundary in the binary image, and an area overlap filter that rejects the smaller ellipse when overlap exceeds $\mathcal{I}_c(x,y)$4 of the smaller ellipse’s area (Abhinav et al., 2017).

The experimental setting comprises two crystal datasets. Dataset-I is heterogeneous and multi-shaped, with two subsets of $\mathcal{I}_c(x,y)$5 images each, about $\mathcal{I}_c(x,y)$6 crystals per image on average, and image size $\mathcal{I}_c(x,y)$7. Dataset-II is homogeneous, with $\mathcal{I}_c(x,y)$8 microscopic images of similarly shaped rectangular crystals, about $\mathcal{I}_c(x,y)$9 crystals per image on average, and the same image size. Compared against SPON and SOEO, the proposed method reports on Dataset-I: Recall $15$00, Precision $15$01, AJSC $15$02, versus SPON Recall $15$03, Precision $15$04, AJSC $15$05, and SOEO Recall $15$06, Precision $15$07, AJSC $15$08. On Dataset-II it reports Recall $15$09, Precision $15$10, AJSC $15$11, versus SPON Recall $15$12, Precision $15$13, AJSC $15$14, and SOEO Recall $15$15, Precision $15$16, AJSC $15$17. Complexity is stated as $15$18, and runtime is about $15$19 s per $15$20 image in Python on a $15$21 GHz CPU with $15$22 GB RAM. For Dataset-I, verification by composition analysis reports an average error of $15$23 in estimated crystal composition (Abhinav et al., 2017).

The relation between these two strands of work is methodological rather than architectural. The 2017 system infers hidden object structure from contour geometry, concavity, and shape-specific fitting; OASIS infers obscured object-specific intensities by weighting the loss toward overlap regions and selected channels. This suggests a broader continuity in overlap-aware segmentation: visible evidence is treated as incomplete, and supervision or geometry is concentrated where overlap obscures the features of interest.