Spatial-Slice Feature Learning++ for CT Analysis

• SSFL++ is an enhanced morphology-guided framework that pre-selects diagnostically useful CT slices by cropping non-lung regions and reducing redundancy.
• The method employs a two-stage pipeline—spatial cropping followed by contiguous slice selection—to retain key lung regions while discarding irrelevant data.
• Augmenting SSFL++ with kernel density-based slice sampling (KDS) significantly boosts performance, making 2D backbones competitive on COVID-19 detection benchmarks.

Searching arXiv for the cited SSFL++ papers to ground the article and verify metadata. Search query: "Spatial-Slice Feature Learning++ COVID-19 Detection arXiv (Hsu et al., 2024)" Spatial-Slice Feature Learning++ (SSFL++) is a morphology-guided framework for chest CT recognition that seeks to reduce redundancy in both the spatial dimension within each slice and the slice dimension across the scan before downstream classification. In the COVID-19 detection setting, it is presented as an enhanced version of earlier Spatial-Slice Feature Learning (SSFL), designed to filter out effectively out-of-distribution slices near the beginning and end of a scan, crop away large non-lung background regions, and retain a compact but representative subset of diagnostically useful slices. In the principal formulations, SSFL++ is paired with Kernel-Density-based slice Sampling (KDS), and the resulting pipeline is used to make simple 2D backbones such as EfficientNet competitive with more explicitly sequential or volumetric models on the COVID-19-CT-DB benchmark (Hsu et al., 2024). Later work treats SSFL++ and KDS as a preprocessing-based input-space standardization scheme for multi-source CT classification, rather than as a new deep architecture in itself (Lee et al., 26 Jul 2025).

1. Definition, naming, and scope

SSFL++ is defined in the CT literature as an enhanced Spatial-Slice Feature Learning framework specifically tailored to variable-length chest CT scans for COVID-19 recognition (Hsu et al., 2024). Its central premise is that raw CT volumes contain substantial redundancy in two orthogonal senses: within a slice, large black background regions and non-lung anatomy can dominate the frame; across the slice axis, many superior and inferior slices contain little meaningful lung tissue for the downstream diagnostic task. The framework therefore aims to locate the relevant anatomical support before feature extraction rather than relying on the classifier to discover it implicitly.

The term is not used uniformly across all related papers. Two 2024 papers explicitly present an “advanced” or “enhanced” SSFL++ framework for COVID-19 CT detection (Hsu et al., 2024). By contrast, the 2023 “Strong Baseline and Bag of Tricks” paper restates SSFL and practical refinements around preprocessing and slice aggregation, but does not introduce a method explicitly called SSFL++ (Hsu et al., 2023). A later multi-source challenge paper says that it “builds upon the SSFL framework” and uses KDS, but also does not formally define SSFL++ as a distinct method (Lee et al., 2 Jul 2025). This suggests that “SSFL++” denotes a family of closely related morphology-driven preprocessing and slice-selection pipelines rather than a single universally standardized algorithmic specification.

A common misconception is to read SSFL++ as a learned attention mechanism or a 3D feature-decomposition network. In the principal COVID-19 papers, it is instead described as an unsupervised, morphology-based preprocessing and selection framework guided by prior knowledge of lung CT structure (Hsu et al., 2024). The learned component comes afterward, typically through a 2D CNN such as EfficientNet-B3, EfficientNet-B7, or related slice-wise backbones.

2. Core pipeline: spatial reduction and slice-range selection

The SSFL++ pipeline has two main stages: a spatial step and a slice step (Hsu et al., 2024). In the spatial step, each slice is first smoothed using a low-pass filter,

$\mathbf{Z}_{\text{filtered}(i,j) = \frac{ \sum_{p=-k}^{k}\sum_{q=-k}^{k} w(p,q)\,\mathbf{Z}(i+p,j+q) }{ \sum_{p=-k}^{k}\sum_{q=-k}^{k} w(p,q) },$

and then thresholded to form a binary mask,

$\mathbf{Mask}[i,j] = \begin{cases} 0, & \text{if } \mathbf{Z}_{\text{filter}[i,j] < t,\ 1, & \text{if } \mathbf{Z}_{\text{filter}[i,j] \ge t. \end{cases}$

The mask determines cropping bounds through the minimum and maximum active coordinates in the two in-plane dimensions. The slice is then cropped to $\mathbf{Z}_{\text{crop}^{c}$ and resized to a fixed input size $H \times W$. The stated purpose of this stage is to avoid direct resizing of full slices with large black margins, which would compress the lung field and reduce effective resolution on diagnostically relevant structures (Hsu et al., 2024).

The slice step is based on a morphology-derived estimate of how much lung tissue is present in each slice. After binary dilation or filling, the method computes

$Area(\mathbf{Z}) = \sum_i \sum_j \mathbf{Mask}_{\text{filled}(i,j) - \mathbf{Mask}(i,j).$

The intended interpretation is that the difference between the filled body mask and the original thresholded mask acts as a proxy for the lung tissue region. Slices with larger $Area(\mathbf{Z})$ are therefore treated as more informative because they correspond more closely to the central lung-bearing portion of the scan (Hsu et al., 2024).

Once per-slice area scores are obtained, SSFL++ selects a contiguous slice interval $[s,e]$ by solving

$\underset{s,e}{\text{maximize} \quad \sum_{i=s}^{e} Area(\mathbf{Z}_i),$

subject to

$e-s \le n_c,$

and

$$\frac{\sum_{i=s}^{e} Area(\mathbf{Z}_i)} {\sum_{i=1}^{n_c} Area(\mathbf{Z}_i)} \ge \alpha.$$

Here, $\mathbf{Mask}[i,j] = \begin{cases} 0, & \text{if } \mathbf{Z}_{\text{filter}[i,j] < t,\ 1, & \text{if } \mathbf{Z}_{\text{filter}[i,j] \ge t. \end{cases}$0 constrains the number of retained slices and $\mathbf{Mask}[i,j] = \begin{cases} 0, & \text{if } \mathbf{Z}_{\text{filter}[i,j] < t,\ 1, & \text{if } \mathbf{Z}_{\text{filter}[i,j] \ge t. \end{cases}$1 specifies the required proportion of total area mass to preserve. The retained interval is meant to be compact but still cover the main lung-bearing region (Hsu et al., 2024). This contiguous-interval formulation distinguishes SSFL++ from arbitrary slice picking and reflects the paper’s claim that global sequence information should be preserved even when most of the scan is discarded.

The redundancy reduction claim follows directly from this two-stage pruning. On the full training and validation data, the average spatial area decreases from $\mathbf{Mask}[i,j] = \begin{cases} 0, & \text{if } \mathbf{Z}_{\text{filter}[i,j] < t,\ 1, & \text{if } \mathbf{Z}_{\text{filter}[i,j] \ge t. \end{cases}$2K to $\mathbf{Mask}[i,j] = \begin{cases} 0, & \text{if } \mathbf{Z}_{\text{filter}[i,j] < t,\ 1, & \text{if } \mathbf{Z}_{\text{filter}[i,j] \ge t. \end{cases}$3K pixels, the average slice length decreases from $\mathbf{Mask}[i,j] = \begin{cases} 0, & \text{if } \mathbf{Z}_{\text{filter}[i,j] < t,\ 1, & \text{if } \mathbf{Z}_{\text{filter}[i,j] \ge t. \end{cases}$4 to $\mathbf{Mask}[i,j] = \begin{cases} 0, & \text{if } \mathbf{Z}_{\text{filter}[i,j] < t,\ 1, & \text{if } \mathbf{Z}_{\text{filter}[i,j] \ge t. \end{cases}$5, and the combined spatial $\mathbf{Mask}[i,j] = \begin{cases} 0, & \text{if } \mathbf{Z}_{\text{filter}[i,j] < t,\ 1, & \text{if } \mathbf{Z}_{\text{filter}[i,j] \ge t. \end{cases}$6 slice volume decreases from $\mathbf{Mask}[i,j] = \begin{cases} 0, & \text{if } \mathbf{Z}_{\text{filter}[i,j] < t,\ 1, & \text{if } \mathbf{Z}_{\text{filter}[i,j] \ge t. \end{cases}$7M to $\mathbf{Mask}[i,j] = \begin{cases} 0, & \text{if } \mathbf{Z}_{\text{filter}[i,j] < t,\ 1, & \text{if } \mathbf{Z}_{\text{filter}[i,j] \ge t. \end{cases}$8M, corresponding to a reduction of $\mathbf{Mask}[i,j] = \begin{cases} 0, & \text{if } \mathbf{Z}_{\text{filter}[i,j] < t,\ 1, & \text{if } \mathbf{Z}_{\text{filter}[i,j] \ge t. \end{cases}$9, or about $\mathbf{Z}_{\text{crop}^{c}$0. On the test set, the total reduction is $\mathbf{Z}_{\text{crop}^{c}$1 (Hsu et al., 2024).

3. Kernel-Density-based slice Sampling

KDS is introduced as the principal addition that distinguishes SSFL++ from earlier SSFL variants using random slice sampling (Hsu et al., 2024). The motivation is that random sampling can be unstable during both training and inference, may miss representative slices, and may overfocus on large-area slices without preserving global coverage of the retained interval. KDS is designed to preserve global sequence information while assigning higher probability to denser, more representative parts of the selected slice range.

The paper describes KDS through kernel density estimation over the selected slice set $\mathbf{Z}_{\text{crop}^{c}$2:

$\mathbf{Z}_{\text{crop}^{c}$3

with Gaussian kernel

$\mathbf{Z}_{\text{crop}^{c}$4

and CDF

$\mathbf{Z}_{\text{crop}^{c}$5

The bandwidth $\mathbf{Z}_{\text{crop}^{c}$6 is chosen by the Scott rule (Hsu et al., 2024). The text says that the probability of selecting slices in each interval is determined by the KDE density while ensuring at least one sample from each sub-interval.

The implementation-level description is only partial. The paper does not explicitly define the final discrete sampling probability function, nor does it fully specify what the sample points $\mathbf{Z}_{\text{crop}^{c}$7 are. The intended algorithmic sequence, however, is explicit: run the SSFL++ slice step to obtain $\mathbf{Z}_{\text{crop}^{c}$8, estimate a smooth density over slice positions in that interval, derive adaptive percentile intervals from the CDF, allocate at least one slice per interval, and then distribute the remaining selections proportionally to density (Hsu et al., 2024). This suggests that KDS is best understood as a density-aware representative sampling rule rather than a learned module.

A later multi-source COVID-19 challenge solution uses the same conceptual machinery in a more deployment-oriented form. It describes KDS as estimating the distribution of slice-wise lung areas and sampling eight representative slices per scan from that distribution, again emphasizing that the method is anatomy-aware rather than explicitly lesion-aware (Lee et al., 2 Jul 2025). This suggests that, in practice, KDS is treated as a fixed-budget sparse sampling method that turns variable-length CT studies into a standardized set of representative 2D inputs.

4. Backbone models, training protocol, and downstream classification

SSFL++ is deliberately paired with simple 2D backbones. The flagship configuration in the 2024 COVID-19 paper is E2D, an EfficientNet-B3-based 2D classifier operating on SSFL++-filtered and KDS-selected slices resized to $\mathbf{Z}_{\text{crop}^{c}$9 (Hsu et al., 2024). The argument is that once redundant regions and slices are removed, strong scan-level performance can be obtained without explicit heavy 3D or spatiotemporal modeling.

The same paper also evaluates E(2+1)D and E3D variants. E(2+1)D first processes SSFL++-filtered scans resized to $H \times W$0, selects 100 slices to encode, converts them into a latent feature queue of size $H \times W$1, then randomly samples 50 features from the queue and applies a simple 1D convolution with kernel size $H \times W$2 along feature dimensions. E3D likewise operates after SSFL++ but uses 3D modeling on sampled slices (Hsu et al., 2024). These comparisons are used to support the claim that a carefully prepared 2D pipeline can be more robust and data-efficient than 2+1D or 3D alternatives in this task.

Training in the main SSFL++ study uses binary cross-entropy, Adam, learning rate $H \times W$3, weight decay $H \times W$4, and batch size 16, with augmentations including HorizontalFlip, RandomScaleShifting, HueSaturationValue, RandomBrightnessContrast, and CoarseDropout (Hsu et al., 2024). The dataset is COVID-19-CT-DB, with 1,684 labeled scans total for training and validation, comprising 873 positive and 811 negative cases; training includes 1,358 scans and validation 326, while a test set of 1,413 scans is used for challenge evaluation (Hsu et al., 2024).

Related papers preserve the same general design philosophy while changing the deployment context. The 2023 SSFL paper uses EfficientNet-B3a for slice-level feature learning, then extracts 224-dimensional embeddings and applies a hybrid 1D-convolution-plus-2D-CNN aggregation pipeline in its “eff-mix-conv” variants (Hsu et al., 2023). The 2025 multi-source challenge solution keeps the slice-based pipeline but compares EfficientNet-B7 and Swin Transformer-Base on eight $H \times W$5 KDS-selected slices per scan (Lee et al., 2 Jul 2025). These studies all present SSFL-style preprocessing as the main mechanism for handling heterogeneous scan lengths and spatial framing, with the backbone architecture treated as secondary.

5. Empirical performance and comparative evidence

The main validation evidence for SSFL++ comes from comparisons on COVID-19-CT-DB. The challenge baseline based on ResNet-50 plus GRU achieves 78.00 scan-level F1 on validation, whereas E2D with SSFL++ preprocessing and KDS reaches substantially higher scores (Hsu et al., 2024). With random sampling, E2D already performs strongly: 8 random slices yield 92.44 slice-level macro-F1 and 93.18 scan-level F1, while 16 random slices yield 92.68 and 93.37, respectively (Hsu et al., 2024). Replacing random sampling with KDS further improves the results: with 8 KDS slices, E2D reaches 93.46 slice-level macro-F1 and 100.00 scan-level F1; with 16 KDS slices, it reaches 94.11 and 100.00 (Hsu et al., 2024).

The ablation study isolates the role of each SSFL++ component. Using E2D as the base, removing everything yields 80.41 macro-F1 at slice level and 81.26 scan-level F1. Adding only the spatial step raises this to 88.01 and 88.04. Adding only the slice step gives 90.32 and 90.48. Combining both, i.e. SSFL++ without KDS, yields 92.68 and 93.37. Adding KDS on top gives 94.11 macro-F1 at slice level and 100.00 scan-level F1 (Hsu et al., 2024). This supports the paper’s interpretation that both spatial cropping and slice-range selection matter, and that KDS contributes an additional gain by making the retained slices more stable and representative.

The hidden test-set results are also favorable. The final E2D-based SSFL++ method reports 94.39 macro-F1, with class-wise F1 of 95.52 for NONCOVID and 93.26 for COVID, compared with the baseline’s 85.11 macro-F1, 87.48 NONCOVID, and 82.74 COVID (Hsu et al., 2024). A related 2025 multi-source challenge report, using SSFL and KDS rather than explicitly naming SSFL++, obtains 94.68 F1 and AUC-ROC $H \times W$6 with EfficientNet-B7 on joint multi-source validation, slightly outperforming Swin Transformer at 93.34 F1 and AUC-ROC $H \times W$7 (Lee et al., 2 Jul 2025).

Later analytical work interprets these gains as architecture-agnostic effects of preprocessing. On a multi-source COVID-19 dataset, baseline EfficientNet-B3 achieves 70.73% F1 and 0.7804 AUC, SSFL++ alone yields 80.49% F1 and 0.8957 AUC, and SSFL++ plus KDS yields 94.68% F1 and 0.9813 AUC. With Swin Transformer, the corresponding values are 68.53% and 0.7654 for baseline, 78.82% and 0.8672 for SSFL++ only, and 93.34% and 0.9797 for SSFL++ plus KDS (Lee et al., 26 Jul 2025). This suggests that the observed improvement is attributed primarily to input standardization rather than to any one backbone.

6. Interpretation, extensions, and limitations

A central interpretation in the later literature is that SSFL++ should be viewed as a preprocessing-based input-space standardization mechanism. In the multi-source CT analysis, SSFL++ is described as the spatial standardization component, enforcing consistent anatomical framing through lung-centric cropping, while KDS is treated as the temporal standardization component, selecting anatomically representative slices from scans with variable length and coverage (Lee et al., 26 Jul 2025). In that view, the combined pipeline reduces inter-source variance before feature learning begins.

This interpretation also addresses a controversy about what SSFL++ actually is. It is not consistently presented as a new end-to-end deep architecture. In the main COVID-19 papers, it is primarily a deterministic pipeline of filtering, thresholding, cropping, contiguous slice-range selection, and density-aware sampling (Hsu et al., 2024). The later domain-shift study explicitly says that it adopts the pipeline from prior work and analyzes its effect, rather than redefining SSFL++ from scratch (Lee et al., 26 Jul 2025). A plausible implication is that SSFL++ is best understood as a structured input representation strategy rather than as a fixed model family.

Several limitations are either explicit or strongly implied. First, the approach depends on morphology-based thresholds and filling operations, so hyperparameters such as $H \times W$8, $H \times W$9, and $Area(\mathbf{Z}) = \sum_i \sum_j \mathbf{Mask}_{\text{filled}(i,j) - \mathbf{Mask}(i,j).$0 may need retuning for different organs, views, datasets, or scanners (Hsu et al., 2024). Second, KDS is only partially specified: the KDE and CDF are described conceptually, but the exact discrete sampling rule is not formalized in full (Hsu et al., 2024). Third, lung-area-based sampling is anatomy-aware rather than lesion-aware; this suggests that small focal abnormalities in atypical slices may not be explicitly prioritized (Lee et al., 2 Jul 2025). Fourth, some reported validation numbers, especially scan-level F1 of 100.00, indicate that more detail on aggregation and variance would be useful (Hsu et al., 2024).

The broader significance of SSFL++ lies in its argument that the main difficulty in CT scan classification is not only backbone design, but also deciding which parts of the scan are worth learning from at all (Hsu et al., 2024). Related SSFL-style work from 2023 reaches a similar conclusion, stating that naive individual-slice prediction followed by merging “lacks slice-wise feature learning” and can decrease performance, whereas important-slice selection and two-stage aggregation improve the outcome (Hsu et al., 2023). In that sense, SSFL++ occupies a distinct methodological niche: it uses hand-crafted anatomical priors and density-aware sampling to make simple 2D deep learning viable for heterogeneous volumetric CT data.