Most medical-imaging detectors read a CT scan as a stack of flat 2D slices rather than as a 3D volume, because processing whole volumes is too memory-hungry. The standard repair, "2.5D" fusion, runs one shared 2D network over a few neighbouring slices and concatenates the results. The neighbouring slices, however, look nearly identical in normal anatomy. That makes their features almost perfectly correlated, which is exactly the setting where unregularized or single-penalty combinations behave badly. We propose to fuse neighbouring slices with an elastic net: a regression penalty that mixes an $L_1$ term, which drives some weights to zero, with an $L_2$ term, which spreads weight evenly across correlated inputs. At every location, the centre slice's features are reconstructed from its neighbours. The fitted combination is the "continuous anatomy" context, and what the neighbours cannot explain, the residual, is a direct signal for focal structures such as nodules that appear in only one or two slices.
The part we expect to matter most is that the two penalty strengths are not tuned by grid search. They are set in closed form at each location. The overall strength follows the estimated image noise, in the spirit of discrepancy-principle rules from variational denoising. The balance between $L_1$ and $L_2$ follows the measured correlation between neighbouring slices, which depends on slice thickness. The central prediction is that thin, highly redundant slices call for ridge-like grouping, while thick, weakly correlated slices call for lasso-like selection. It follows that a single fixed mixture is wrong across scanners, and that adapting it should remove much of the known accuracy drop of 2.5D detectors on thick-slice scans.
Core Claim
The claim is that 2.5D slice fusion is a collinear regression problem whose right regularizer changes with two measurable scan properties, noise level and slice spacing. A two-parameter elastic net whose parameters are computed from those properties at each location should match or beat fixed fusion schemes at every slice thickness and noise level. It should do so without per-dataset hyperparameter search, and should cut false positives on structures that continue across slices, such as vessels.
This combines the three directions in unequal roles. Volumetric context loss is the application and the failure we want to fix. The elastic net is the mechanism, chosen because its grouping behaviour on correlated inputs matches the physics of adjacent slices. Automated regularization tuning supplies the method for setting the penalty strengths, and it is extended here to the case its literature leaves open: more than one penalty, varying over space.
Background and Positioning
Three lines of 2.5D and slice-based work motivate the problem. 3DCE runs a shared 2D backbone over 3 to 27 neighbouring slices and concatenates the feature maps before a detection head. It gains 18 to 54 mm of context at 2D cost, but it also reports that accuracy degrades when slices are 5 mm instead of 1 mm, because the concatenation treats all neighbours alike regardless of how far apart they are physically. Roth et al. show that independent 2D predictions give jagged pancreas boundaries and repair them with 3D Gaussian smoothing of probabilities. That is a fixed, isotropic smoother applied after the fact. CARVE is the closest in spirit. It argues that volumes are anisotropic, with heavy redundancy between adjacent slices and diagnostically crucial lesions spanning only one or two, and it selects tokens by high local feature residuals for efficient vision-language processing.
Our proposal differs from all three in the same way: the context weights and the residual come out of an explicitly regularized regression whose penalties are set by scan properties. They are not produced by concatenation, fixed smoothing, or a heuristic selection score. CARVE also targets token budgets for vision-language models, whereas we target detection accuracy and robustness to slice spacing.
The tuning rule borrows from the automated regularization literature, which also constrains the design. Caron et al. adjust $\lambda$ with an outer Uzawa loop until residual energy matches the noise variance. Total Directional Variation for Video Denoising derives closed-form scaling of its parameters with estimated noise standard deviation. Wang et al. show that static parameters fail across noise levels $\sigma \in \{15, 25, 35, 50\}$ while an adaptive search holds up.
We take two things from this literature. The first is the discrepancy idea, that penalty strength should scale with expected residual energy. The second is the preference for closed-form rules, because the research file notes that outer loops add latency that is unacceptable in clinical pipelines. The same literature, as summarized in the file, is limited to single scalar penalties and to global or per-group parameters. Our contribution to it is a rule for a two-parameter composite penalty that varies with location.
We also take a caution from SHASTA-PCA and Bayesian topology learning. Joint noise estimation there relies on a parametric noise model, and CT noise is not Gaussian. We therefore make the rule's constants learnable instead of trusting a derived formula blindly.
On the elastic net side, the file notes that pure lasso picks one variable arbitrarily out of a correlated cluster and becomes unstable (Generalized LARS). For 2.5D fusion, that means a lasso-only fusion would flicker between neighbouring slices from one location to the next. This failure is testable, and we include it as an ablation.
The intersection section of the research file was not available to us, so we judge novelty only against the papers listed above.
Mechanism
A shared 2D backbone produces feature maps for a centre slice $c$ and $K$ neighbours. At each spatial location, let $f_c \in \mathbb{R}^C$ be the centre feature vector and $F \in \mathbb{R}^{C \times K}$ the stacked neighbour features. We solve
$$ w^\star = \arg\min_w \; \|f_c - F w\|_2^2 + \lambda \big( \alpha \|w\|_1 + (1-\alpha)\|w\|_2^2 \big), $$
where $w$ holds the $K$ neighbour weights, $\lambda$ is the overall penalty strength, and $\alpha \in [0,1]$ sets the share of the sparsity-inducing term. The detection head receives the centre feature $f_c$, the context summary $Fw^\star$, and the residual $r = f_c - Fw^\star$. A vessel that continues across slices is well explained by its neighbours and leaves a small residual. A nodule confined to one slice leaves a large one.
The solver is a small fixed number of unrolled proximal-gradient steps. $K$ is at most around 9 per side, so each location's problem is tiny and the whole operation is differentiable.
The parameters are set per location and per scan by $\lambda = a\,\hat\sigma^2$ and $\alpha = \mathrm{sigmoid}(b_0 - b_1 \hat\rho)$. Here $\hat\sigma^2$ is the image noise variance, estimated from high-frequency content in a local window. $\hat\rho$ is the mean correlation between the centre feature and its neighbours in that window. The scalars $a, b_0, b_1$ are learned end to end.
The functional forms come from two sources. Scaling $\lambda$ with $\sigma^2$ follows the discrepancy principle, because expected residual energy grows with noise variance. Letting $\alpha$ fall as correlation rises follows the elastic net's grouping behaviour, which only helps when inputs are correlated. There are only three learned constants, which keeps the rule interpretable and makes it checkable against a grid-searched oracle.
Research Program
The work has three parts. First, we implement the fusion layer as a drop-in replacement for 3DCE's concatenation, keeping the backbone and detection head fixed, so that differences come only from fusion. Second, we build controlled test conditions. Thick-slice versions of thin-slice scans are made by discarding every second, third or fourth slice, which gives the same patients and lesions at different spacings. Noise-perturbed versions are made by injecting noise at graded levels, borrowing the graded-$\sigma$ design of Wang et al. Third, we analyze what the rule learns. The main question is whether the oracle-best $\alpha$ found by grid search at each thickness actually moves monotonically with measured slice correlation. If it does not, the core premise is wrong and the adaptive rule has nothing to track.
Baselines
The naive reference point is a single-slice 2D detector with the same backbone and head and no neighbouring context. It shows how much any volumetric context is worth at each slice thickness.
The strongest existing method is 3DCE's shared-backbone concatenation (3DCE). It is the published 2.5D design on DeepLesion and the one whose thickness sensitivity we claim to fix. It will be re-run with our backbone, the same neighbour count $K$, the same training schedule and the same augmentation, not quoted. We found no reported result on Lacuna for 3DCE's DeepLesion sensitivity figures, so every comparison uses our own reproduction. Beating it means higher sensitivity at a fixed low false-positive rate on thick-slice subsets, with no loss on thin slices.
A stronger modern competitor is learned attention over slices: a softmax-weighted sum of neighbour features, with the weights produced by a small network. It is given the same inputs, including $\hat\sigma$ and slice spacing as extra inputs, and the same training budget.
The grid-tuned oracle is the same elastic net layer with $(\lambda, \alpha)$ fixed per thickness bucket and chosen by grid search on validation data. Its search cost is counted in GPU-hours. This baseline isolates whether adaptivity beats well-chosen constants. Matching it at zero search cost is the minimum bar; beating it requires within-scan spatial variation to matter.
Two ablations isolate the key ingredient. The first adapts $\lambda$ from noise but fixes $\alpha$, which mirrors single-parameter noise-adaptive tuning from prior work. The second uses pure lasso and pure ridge self-expression. If adapting $\lambda$ alone captures most of the gain, the elastic net mixture is not pulling its weight.
Why not just…?
Why not just use learned attention over slices, with slice thickness as an input? This is the most serious threat, because a flexible network could in principle learn whatever weighting the elastic net rule computes. Our argument is about data efficiency and robustness. The rule has three constants, while attention must learn its thickness dependence from whatever thickness mix exists in training. The file notes that native 3D models already overfit under 3D data scarcity. We test this directly by training both methods on thin-slice data only and evaluating on thick-slice subsets. If attention extrapolates as well, the proposal's claim shrinks to interpretability, and we will report it that way.
Why not just grid-search $\lambda$ and $\alpha$ per scanner protocol? That is the traditional answer the automated tuning literature tries to replace. It may well recover most of the thickness gain, and the oracle baseline measures exactly that. What a per-protocol constant cannot capture is variation within a scan, for example noise that differs between shoulders and abdomen. We therefore also report results stratified by local noise level within scans. If the oracle ties everywhere, the contribution is removing the search, not improving accuracy.
Why not just resample volumes to isotropic spacing, or use CARVE-style residual selection? Interpolating thick slices creates correlated neighbours that carry no new information. We expect it to inflate $\hat\rho$ and mislead any fusion method, and we test this as an extra baseline. CARVE's residual scoring (CARVE) is a selection heuristic built for token budgets. Plugging its residual into our head, in place of the elastic net residual, tests whether regularization matters beyond simply computing a difference from neighbours.
Proposed Experiments
The primary benchmark is DeepLesion, evaluated with free-response curves: sensitivity at 0.5, 1, 2 and 4 false positives per image. We break results down by native and subsampled slice thickness and by injected noise level. A secondary check uses the pancreas segmentation setting of Roth et al.. There we measure boundary consistency between adjacent slices against their 3D smoothing and against 3DCE-style fusion.
Four checks will decide the outcome. First, the thin-versus-thick sensitivity gap should shrink relative to reproduced 3DCE. Second, the grid-searched best $\alpha$ should track measured correlation monotonically. This is the premise test, and failure here ends the thesis. Third, sensitivity should stay flat under increasing injected noise with the adaptive $\lambda$, while fixed-$\lambda$ fusion degrades. Fourth, false positives on vessel-like structures should drop. We measure this on a radiologist-reviewed sample of false-positive detections, checking whether the residual channel suppresses structures that continue across slices.
References
- 3D Context Enhanced Region-Based Convolutional Neural Network for End-to-End Lesion Detection
- Deep convolutional networks for pancreas segmentation in CT imaging
- CARVE: Cross-Slice Anisotropic Reallocation of Visual Evidence for Efficient 3D Medical Volume Understanding
- Deblurring of irregularly sampled images by TV regularization in a spline space
- Total Directional Variation for Video Denoising
- Image Denoising Based on TVD Runge-Kutta Method And Projection Algorithm
- Streaming Heteroscedastic Probabilistic PCA with Missing Data
- Bayesian Topology Learning and noise removal from network data
- Generalized LARS as an effective feature selection tool for text classification with SVMs
- DeepLesion: automated mining of large-scale lesion annotations and universal lesion detection with deep learning