DeepAf: One-Shot Spatiospectral Autofocus

• DeepAf is a deep-learning based autofocus model that infers the optimal focal distance from one blurry image using a hybrid spatiospectral approach.
• It integrates a U-Net-like spatial encoder with Fast Fourier Convolution blocks to capture both tissue morphology and defocus physics, enhancing cross-protocol robustness.
• The framework reduces scanning time by 80% while preserving diagnostic quality, making automated slide digitization feasible in resource-constrained settings.

DeepAf is a single-shot, deep-learning-based autofocus framework for digital pathology that predicts the distance to the optimal focal plane from a single input image by combining spatial and spectral information. Introduced in “DeepAf: One-Shot Spatiospectral Auto-Focus Model for Digital Pathology” (Yeganeh et al., 6 Oct 2025), it is positioned as an alternative to Whole Slide Imaging (WSI) scanners and to autofocus pipelines based on focal stacks, with the explicit aim of enabling automated slide digitization on conventional microscopes in resource-constrained settings. The framework is presented not only as a neural model, but as part of a complete hardware-software system that performs tissue detection, single-shot autofocus, and image acquisition while preserving downstream diagnostic utility.

1. Problem setting and motivation

In digital pathology, WSI scanners are described as the gold standard for digitizing pathology samples, but their high cost limits accessibility in many healthcare settings. Lower-cost microscope-based digitization faces a central technical difficulty: histology sections retain 3D morphology, so the optimal focal plane varies across the slide. If a region is acquired out of focus, image quality degrades and downstream diagnosis or AI analysis can be affected (Yeganeh et al., 6 Oct 2025).

DeepAf targets this autofocus bottleneck. Traditional autofocus methods are reported to require full focal stacks, repeated focus measurements at many z-positions, or heuristic optimization, and are therefore slow. Existing learning-based methods are described as either requiring multiple input images or lacking generalization across tissue types and staining protocols. DeepAf is proposed specifically to infer the distance to the optimal focal plane from only one blurry image, thereby reducing scan time, reducing data acquisition overhead, and enabling real-time operation on low-cost hardware.

The method is motivated by a physical prior about defocus in the frequency domain: out-of-focus images lose high-frequency content and exhibit altered spectral cut-off patterns. This motivates the explicit use of a spatiospectral architecture rather than a purely image-space model. A common misconception would be to view DeepAf as only a spatial CNN for blur estimation; the paper instead defines it as a hybrid framework in which frequency-domain information is a first-class component of the focus prediction pipeline.

2. Spatiospectral formulation and network design

DeepAf is formulated as a regression model

$$f: \mathbb{R}^{H \times W \times C} \rightarrow \mathbb{R},$$

with the specific composition

$f(I) = R(B(E_s(I), E_\omega(I))).$

Here, $I \in \mathbb{R}^{H \times W \times C}$ is the input image, $E_s$ is the spatial encoder, $E_\omega$ is the spectral encoder, $B$ is the bottleneck or fusion layer, and $R$ is the regression head (Yeganeh et al., 6 Oct 2025).

The spatial encoder $E_s$ is a U-Net-like encoder with four hierarchical convolutional layers. It extracts local tissue morphology, texture, and structural cues, and maps the input as

$$E_s: \mathbb{R}^{H \times W \times C} \rightarrow \mathbb{R}^{h \times w \times d_s}.$$

The spectral encoder $E_\omega$ uses four sequential FFC blocks (Fast Fourier Convolution blocks) and maps

$f(I) = R(B(E_s(I), E_\omega(I))).$0

The two feature streams are concatenated,

$f(I) = R(B(E_s(I), E_\omega(I))).$1

processed by the bottleneck $f(I) = R(B(E_s(I), E_\omega(I))).$2, and then reduced by a regression head consisting of 2D average pooling and a linear layer. The scalar output is the predicted focus offset $f(I) = R(B(E_s(I), E_\omega(I))).$3, i.e. the distance to the optimal focal point along the microscope’s $f(I) = R(B(E_s(I), E_\omega(I))).$4-axis.

The spectral branch is justified by the paper’s characterization of focus in the power spectrum: out-of-focus images attenuate low-frequency components near the center of the power spectrum, whereas in-focus images have stronger low-frequency amplification and a higher cut-off frequency. This suggests that the spectral encoder is intended to encode defocus physics directly rather than only statistical image regularities.

3. Regression objective and autofocus control loop

DeepAf is trained as a continuous regression system rather than a classifier. Given an image patch, it predicts the continuous defocus amount and is trained with smooth $f(I) = R(B(E_s(I), E_\omega(I))).$5 loss, denoted in the paper as

$f(I) = R(B(E_s(I), E_\omega(I))).$6

The predicted $f(I) = R(B(E_s(I), E_\omega(I))).$7 is then used by the motor controller to move the focus stage to the correct z-position (Yeganeh et al., 6 Oct 2025).

The framework is embedded in an automated microscopy system built around a SWIFT-380t microscope. The reported hardware includes a motorized stage with three stepper motors, $f(I) = R(B(E_s(I), E_\omega(I))).$8 and $f(I) = R(B(E_s(I), E_\omega(I))).$9 for x-y slide motion and $I \in \mathbb{R}^{H \times W \times C}$0 for z-axis focus control, with focus precision of 0.002 mm. A camera mounted above the objective lens captures low-resolution images for tissue detection and high-resolution images after focusing. A Raspberry Pi control system coordinates motor movement and image acquisition and runs DeepAf inference on CPU.

The scan pipeline is described as four stages: slide scanning along a predefined trajectory, tissue detection using HSV thresholding, single-shot autofocus using DeepAf, and high-resolution capture at the predicted focus $I \in \mathbb{R}^{H \times W \times C}$1. Tissue detection uses the criterion

$I \in \mathbb{R}^{H \times W \times C}$2

where $I \in \mathbb{R}^{H \times W \times C}$3 are pixel values in the HSV Value channel and $I \in \mathbb{R}^{H \times W \times C}$4 is the threshold. If the region is non-empty, the image is passed to DeepAf, the system adjusts the z-stage accordingly, and the final image is acquired.

This integrated design is significant because the paper does not present autofocus as an isolated inference task. Instead, focus prediction is coupled to stage control and scanning throughput, so the value of the model is inseparable from its deployment within an automated acquisition loop.

4. Experimental design and data regimes

The paper evaluates DeepAf in two settings: an incoherent autofocus benchmark and a brain tissue case study (Yeganeh et al., 6 Oct 2025).

For the autofocus benchmark, the dataset from Jiang et al. is used with same-lab and different-lab test sets at 20x magnification. The depth of field is 1 µm, and patches are $I \in \mathbb{R}^{H \times W \times C}$5. For the brain tissue prototype, 406 brain tissue samples are used for autofocus training, and 536 histopathology slides are scanned for the clinical study. The tissue categories are high-grade glioma, low-grade glioma, inflammatory, and normal. This case study operates at 4x magnification with a depth of field of 60 µm, and focal stacks of 1000 slices per sample are created during training.

Training details for the brain tissue study are reported as approximately 130K patches, an 80/20 train-validation split, 700 test patches, batch size 32, learning rate $I \in \mathbb{R}^{H \times W \times C}$6, weight decay 0.006, 100 epochs, and Adam optimizer. Data augmentation includes channel-wise normalization, random erasing, Gaussian blur, random perspective, random auto contrast, and color jittering.

Evaluation uses focus error (FE), defined as the mean absolute error between predicted and optimal focal distance, reported with standard deviation. In addition to FE, the paper reports false focus prediction rate and the percentage of predictions within the depth of field. These latter measures are especially relevant because they characterize operational failure modes: whether the prediction lands on the wrong side of the slide and whether the resulting image would still appear acceptably sharp.

5. Autofocus performance, generalization, and ablations

DeepAf is compared against prior autofocus methods on same-protocol and different-protocol test sets. The reported focus errors are summarized below (Yeganeh et al., 6 Oct 2025).

Setting	Method	Focus error
Same-protocol	Dastidar et al. (2 images)	0.19 µm
Same-protocol	Jiang et al. (1 image)	0.46 µm
Same-protocol	Chen et al. (1 image)	0.21 µm
Same-protocol	DeepAf Spatial	0.18 µm
Same-protocol	DeepAf Spatiospectral	0.18 µm
Different-protocol	Dastidar et al. (2 images)	0.25 µm
Different-protocol	Jiang et al. (1 image)	0.53 µm
Different-protocol	Chen et al. (1 image)	0.44 µm
Different-protocol	DeepAf Spatial	0.39 µm
Different-protocol	DeepAf Spatiospectral	0.32 µm

On same-protocol data, DeepAf Spatial and DeepAf Spatiospectral both report 0.18 µm FE, matching or exceeding prior methods while using only one input image. On different-protocol data, the spatiospectral variant improves from 0.39 µm to 0.32 µm relative to the spatial-only variant, while outperforming the cited single-image baselines. The paper also reports model sizes of 4.7M parameters for DeepAf Spatial, 4.2M parameters for DeepAf Spatiospectral, 10.8M parameters for Jiang et al., and 3.5M parameters for Dastidar et al.

The cross-lab generalization metrics emphasized for the spatiospectral model are a false focus prediction rate of 0.72% and approximately 90% of predictions within the depth of field. The ablation table further separates spatial, spectral, and spatiospectral variants. At the same 20x magnification, the reported metrics are: Spatial, false focus prediction 0.86%, DoF 89.24%, FE 0.18; Spectral, false focus prediction 1.29%, DoF 73.03%, FE 0.29; Spatiospectral, false focus prediction 0.72%, DoF 89.81%, FE 0.18. Under different protocol at 20x magnification, the reported metrics are: Spatial, false focus prediction 1.22%, DoF 71.34%, FE 0.39; Spectral, false focus prediction 3.12%, DoF 53.81%, FE 0.46; Spatiospectral, false focus prediction 1.60%, DoF 73.78%, FE 0.32.

These results support two narrower interpretations stated or implied by the paper. First, the spectral branch appears most valuable for cross-protocol robustness rather than for single-lab optimization. Second, the hybrid model balances the strengths of spatial and spectral cues, even though on the authors’ own 4x dataset the purely spatial model slightly outperformed it. The paper therefore does not claim uniform superiority of the spectral component in every regime.

6. Throughput, clinical study, and significance for digital pathology

Operationally, DeepAf is reported to reduce focusing time by 80% compared to stack-based methods (Yeganeh et al., 6 Oct 2025). Because autofocus is identified as one of the most time-consuming parts of scanning, this reduction directly affects slide throughput and the practicality of automated microscopy.

The paper extends evaluation beyond autofocus error to a downstream clinical task using images acquired by the low-cost 4x prototype. On 536 brain tissue slides, binary classification is defined as cancer = high-grade glioma + low-grade glioma and normal = inflammatory + normal. The reported result is AUC = 0.90 and F1 score = 0.83, evaluated with 5-fold cross-validation. The paper emphasizes that this is achieved at 4x magnification, which is substantially lower than typical 20x WSI workflows.

This downstream result matters because it links focusing performance to diagnostic utility. A plausible implication is that autofocus quality cannot be evaluated solely through FE: the clinically relevant question is whether the acquired imagery remains useful for pathology inference. The reported AUC of 0.90 indicates that the resulting images retained sufficient information for meaningful cancer classification in the study setting.

The broader significance assigned to DeepAf is therefore not limited to autofocus accuracy. The framework is presented as a route toward accessible digital pathology, automated slide scanning using conventional microscopes, and lower-cost deployment in resource-constrained settings. The paper frames this as a hardware-software co-design for real-time digital pathology rather than as an isolated computer-vision model.

7. Limitations, interpretation, and relation to prior autofocus assumptions

The paper notes or implies several limitations that constrain interpretation of the results (Yeganeh et al., 6 Oct 2025). Performance varies by protocol: the spectral branch generalizes better across protocols, but the spatial-only branch can be stronger on single-lab data. The curated brain dataset is single-lab, so the generalization of the individual model variants cannot be fully assessed there. The clinical validation is performed at 4x magnification, which the paper identifies as lower than typical 20x WSI scans and therefore not directly comparable to conventional WSI-based studies. Finally, the model still exhibits a nonzero false-focus prediction rate, so autofocus is not error-free.

These caveats are important when distinguishing what the paper demonstrates from what it suggests. It demonstrates one-shot focus regression from a single image, integration into a motorized microscope system, 0.18 µm same-protocol FE, 0.32 µm different-protocol FE, 0.72% false focus predictions for the spatiospectral model, approximately 90% predictions within the depth of field, and an 80% reduction in focusing time. It also demonstrates downstream classification performance of 0.90 AUC at 4x magnification on the reported brain tissue study. It does not claim replacement of WSI scanners in all settings, nor does it show that the spatiospectral variant dominates the spatial variant on every dataset.

In methodological terms, DeepAf occupies a distinct position within computational microscopy autofocus. Traditional autofocus stacks infer focus by exhaustive sampling in z; many prior learning methods use multiple views or images; DeepAf instead treats autofocus as direct continuous regression from a single image, with explicit fusion of spatial morphology and spectral defocus signatures. That combination is the defining characteristic of the framework and the main basis for its reported cross-lab robustness.