{"id":"a05cd250-302c-4d7d-a9e8-339230626a40","arxiv_id":"2505.09178","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A frozen ViT with 0.17% trainable LoRA parameters and a unified 2D/3D embedding layer matches or slightly beats fully fine-tuned ViTs across 12 medical diagnosis datasets.","lead":"UniCAD is a multi-task medical imaging architecture that keeps a pre-trained vision transformer frozen and adds tiny trainable LoRA adapters per task, reporting accuracy on par with full fine-tuning while using only 0.17% trainable parameters. It also processes both 2D and 3D images through one system, which could simplify clinical deployment and enable sharing of lightweight diagnostic experts.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 0.17% parameter claim conflicts with the memory table's DEAR logic and is never defined precisely, so the paper's headline efficiency claim is not reliably supported.","rationale":"The core empirical claim, that LoRA experts on a frozen ViT roughly match fully fine-tuned ViT, is plausible and largely supported by Tables 1 and 2: UniCAD is at or near the top on most tasks. The reader's weakest_assumption focuses on the 3D transfer of 2D attention priors, which is a genuine technical risk, but I find a more specific and load-bearing problem in the efficiency metric itself. The headline 0.17% number is undefined and inconsistent with the DEAR table's presentation, and the DEAR metric's arbitrary exponent k=3 and aggregate memory term make the efficiency comparison hard to interpret. These issues do not invalidate the accuracy comparisons, but they do undermine the precision of the central efficiency claim. A focused re-derivation of DEAR and the parameter counts is the most direct way to settle whether the paper's main quantitative contribution holds.","tokens_in":19036,"tokens_out":1245,"duration_ms":13238,"concrete_test":"Recompute Table 3's DEAR values from Eq. (3) with k=1 and with the memory term computed per-method per-task rather than as a single total; then recompute the 0.17% figure from the released checkpoints and the stated LoRA ranks. If the DEAR advantage over fine-tuned ViT collapses to near equality, or if 0.17% is not reproducible, the central efficiency claim is not supported as stated.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The abstract and highlights state that UniCAD introduces only 0.17% trainable parameters, and the Introduction generalizes this to '0.17% of total weights.' Yet the paper never defines the denominator: 0.17% of which model (ViT-base? ViT-giga?), and does it include the task-specific classification heads? Table 3 reports UniCAD at 0.93G GRAM for 11 tasks versus 4.48G for fully fine-tuned ViT, but the DEAR metric in Eq. (3) uses an exponent k=3 and a memory ratio in the denominator, making the five-fold DEAR gap (1.010 vs 0.204) driven mostly by memory and the arbitrary exponent, not by accuracy. The paper also does not state LoRA rank r, patch sizes, or how the 3D positional embeddings are initialized relative to the 2D pretrained ones (Sec. 3.2, 4.2), which means the 0.17% figure cannot be reproduced from the text. A metric that is central to the paper's quantitative contribution should be tightly specified and consistently used.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes UniCAD, a multi-task computer-aided diagnosis (CAD) architecture that combines a frozen, pretrained Vision Transformer (ViT) backbone with a trainable Unified Embedding Layer (UEL) to handle both 2D and 3D medical images, and per-task low-rank adaptation (LoRA) experts. The claimed contributions are (1) parameter efficiency—only 0.17% trainable parameters relative to full fine-tuning—paired with accuracy comparable to fully fine-tuned ViTs across 12 diagnostic datasets; (2) plug-and-play extensibility via lightweight task-specific experts; and (3) an open ecosystem for sharing CAD experts. The paper introduces a new evaluation metric, DEAR, to capture the accuracy-memory tradeoff in multi-task deployment, and reports experiments comparing UniCAD with foundation models (CLIP variants, Med-Flamingo, M3D-LaMed), fine-tuned ViT/ResNet, and linear probing.","tokens_in":19210,"tokens_out":9468,"duration_ms":85277,"significance":"If the efficiency claims are validated, the work addresses a practically important problem: deploying multiple diagnostic tasks on a shared vision foundation model with minimal GPU memory and trainable-parameter overhead. The benchmark coverage is reasonably broad (8 2D and 4 3D datasets), and the architecture is simple and plausible. The authors provide source code and a project page, which supports reproducibility. The accuracy results appear to be independent measurements, and the paper is honest about the limitation that expert training still requires annotated data. However, the central efficiency metric (DEAR) is not reproducible from the reported numbers, the headline 0.17% parameter figure is underspecified, and the 3D transfer assumption is not ablated. These issues currently prevent a reader from verifying the paper's primary quantitative claims, although the core architecture remains a credible direction for future work.","major_comments":[{"comment":"The DEAR values in Table 3 are inconsistent with the definition in Eq. (3). Using the reported Avg ACC values for the 11 tasks, the numerator is (0.795/0.788)^3 ≈ 1.03. For UniCAD's reported DEAR of 1.010, the denominator would need to be ≈ 1.02, implying an average per-task memory for the ViT baseline of about 0.083 GB—far below the 1.45 GB reported for a single fine-tuned ViT in Table 4. If instead the GRAM totals in Table 3 are used directly (0.93 GB for UniCAD vs 4.48 GB for ViT), the denominator under any natural reading of Eq. (3) gives DEAR values of roughly 4.9 or 54.5, not 1.010. Please specify precisely what E_i^M and E_ViT represent (per-task memory, total system memory, or something else) and provide a step-by-step numerical example for one row of Table 3. Without this, the paper's headline efficiency metric cannot be reproduced or interpreted.","section":"§4.3, Eq. (3), Table 3"},{"comment":"The claim that UniCAD introduces only 0.17% trainable parameters is not precisely defined or reproducible. The LoRA rank r is never stated, and the paper does not specify whether the 0.17% is relative to the ViT-base backbone only, includes the classification heads, or uses some other denominator. Section 3.5 reports only the learning rate, optimizer, and training epochs; it omits the LoRA rank, patch sizes for 2D and 3D inputs, and input resolutions. These are all load-bearing hyperparameters: without them, the 0.17% figure and the parameter-efficiency comparisons in Table 3 and Figure 7 cannot be verified. Please state the exact hyperparameters and a formula for the parameter count.","section":"Abstract, §1, §3.3, §3.5"},{"comment":"The central assumption that a frozen, 2D-pretrained ViT can reason effectively about 3D patch tokens is not tested. The paper does not state how the 3D positional embeddings are initialized relative to the pretrained 2D ones (interpolation, learned from scratch, or something else), and it does not provide an ablation that varies the 3D embedding strategy or compares against a 3D-pretrained backbone. This matters because the 3D accuracy differences between UniCAD and fully fine-tuned 3D ViT are very small (e.g., 0.521 vs 0.517 on FractureMNIST3D, 0.848 vs 0.842 on NoduleMNIST3D) and could easily be within run-to-run variation. The claim of \"performance on par\" on 3D tasks requires error bars or statistical significance testing, and the transfer assumption needs a direct ablation.","section":"§3.2, §4.2, Table 1"},{"comment":"The memory measurements for fine-tuned ViT in Table 4 appear inconsistent with the narrative that all task-specific ViTs must be loaded simultaneously. For a single task the reported GRAM is 1.45 GB (base model, batch size 1), but for 8 tasks it is only 3.94 GB, which is sublinear and contradicts the claim that each extra task adds a full ViT. This also conflicts with Table 3, where 11 fine-tuned ViT tasks are listed as 4.48 GB total. Please clarify what is actually measured (e.g., shared backbone weights, sequential processing, or only a subset of models loaded) and ensure that memory definitions are consistent across Tables 3 and 4, because these numbers directly feed into the DEAR metric and the deployment-efficiency conclusions.","section":"Table 4"}],"minor_comments":[{"comment":"For reproducibility, please report the batch size, weight decay, LoRA dropout, data augmentation, and the exact optimizer settings, in addition to the learning rate and epoch count.","section":"§3.5"},{"comment":"The column layout of Table 1 is hard to read because the model category headers span across both 2D and 3D blocks. Consider splitting the table into two separate panels or formatting the header rows so that each model's category is clearly associated with its columns.","section":"Table 1"},{"comment":"The exponent k=3 is described only as \"empirically introduced.\" Please provide a sensitivity analysis showing how the DEAR rankings change for different values of k (e.g., 1, 2, 3, 4) so that the metric's conclusions are not driven by an arbitrary choice.","section":"§4.3"},{"comment":"Figure 7 plots accuracy versus foundation model scale, but the exact parameter counts and the LoRA rank used at each scale are not reported in the text. Adding these details would let readers replicate the scaling curves.","section":"Figure 7"},{"comment":"The reference to Hu et al. (LoRA) is missing the publication venue and year; the entries for MoELoRA and LViT should also be checked for completeness.","section":"References"},{"comment":"The symbol PMC-CLIP* is used in the tables and text but is only partially defined earlier; please state explicitly in the table footnotes and in §4.2 that the asterisk denotes the fine-tuned version of PMC-CLIP.","section":"§4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper appears to build closely on the authors' MeLo work (Zhu et al., 2024); the editorial board may want to verify the novelty overlap, especially since the UEL and the multi-task batching scheme are the main technical additions. The efficiency claims are the primary selling point of the manuscript, so the lack of a reproducible DEAR metric is a serious issue that will require a full re-computation and possibly a reformulated metric. The 0.17% parameter claim and the unstated LoRA rank/patch sizes also weaken the paper's core quantitative message. The accuracy measurements themselves seem to be independent, which is a positive sign; I do not see evidence of circular reasoning in the main results. The 'open ecosystem' section is largely aspirational and might be better framed as a future-work or platform announcement rather than a concrete technical contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read UniCAD. Quick take: the accuracy results are credible; the efficiency metric is not in its current form. The Unified Embedding Layer (UEL) is a real, though modest, extension of the authors' earlier MeLo work—it lets a frozen 2D-pretrained ViT consume 3D patches via a trainable projection and separate positional embedding, and the rank-standardized batching is a sensible engineering fix. The evaluation across 12 datasets is broader than what MeLo or MOELoRA had, and the numbers are in line with what I'd expect from LoRA: on par with full fine-tuning on nearly every task. What's good: the paper is upfront in Sec. 1 that it builds on MeLo. The 0.17% trainable-parameter figure is plausible for LoRA on query/value projections, and the throughput simulation with mixed-task batching shows a genuine deployment benefit. The scaling experiment up to ViT-giga is a nice addition. Soft spots: the DEAR metric in Eq. 3 does not reproduce the values in Table 3. I recomputed UniCAD: accuracy ratio ~1.009, cubed 1.027; memory ratio 0.93/4.48 = 0.208; quotient ~4.9, not the reported 1.010. For fine-tuned ViT, DEAR should be 1.0 but is listed as 0.204. The stress-test note is right: the 0.17% figure is never precisely defined—LoRA rank r is not stated, patch sizes appear only symbolically, and the initialization of the 3D positional embeddings is not described. No error bars either. These are fixable but currently block reproduction. Also, MeLo and MOELoRA are cited but not compared experimentally; that is needed to see what UEL adds beyond its predecessors. None of this sinks the central empirical claim that a frozen ViT plus LoRA experts matches full fine-tuning across these tasks; that is consistent with prior work. But the DEAR inconsistency undercuts an explicitly claimed contribution, and the missing details are a completeness problem that makes the headline 0.17% hard to verify. Who it's for: anyone working on parameter-efficient, multi-task medical image diagnosis. It deserves a serious referee; I would send it to peer review, but the authors need to specify the architecture precisely, fix or properly explain DEAR, and add the missing baseline comparisons.","headline":"UniCAD's accuracy story is believable and its 2D/3D unified embedding is a useful extension of MeLo, but the DEAR efficiency metric is internally inconsistent and key implementation details are missing.","tokens_in":759,"tokens_out":1085,"would_cite":true,"duration_ms":56763,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A frozen vision transformer with 0.17% trainable parameters matches fully fine-tuned models across 12 medical diagnosis tasks.","keywords":["computer-aided diagnosis","foundation model","low-rank adaptation","vision transformer","medical image classification","2D/3D unified embedding","parameter-efficient fine-tuning","clinical deployment"],"falsifier":"Run the four 3D tasks twice, initializing the UEL's 3D patch embedding and the 3D positional embeddings once by inflation or interpolation from the pretrained 2D weights and once from random values: if accuracy on NoduleMNIST3D and AdrenalMNIST3D collapses in the second case, the frozen attention's 3D transfer is load-bearing; if both runs match, the trainable embedding alone carries the result. A complementary check is a 3D dataset with strong orientation sensitivity, where a 2D attention prior might plausibly fail, which would mark the scope of the claim.","tokens_in":18808,"feed_emoji":"🩻","tokens_out":12696,"duration_ms":100900,"temperature":0.7,"pith_summary":"UniCAD claims that one frozen vision transformer can serve as the common engine for many medical diagnosis tasks at once, matching the accuracy of fully fine-tuned per-task models while training only 0.17% of the weights. Each task contributes a pair of low-rank matrices injected into the attention layers plus a small classification head, so adding a task means exchanging a lightweight module rather than retraining a model. The paper reports this recipe working across 12 datasets spanning 2D X-rays, histology, retina, mammography, and dermatoscopy images as well as 3D CT and MRI volumes, and shows accuracy rising as the frozen backbone scales from 82 million to 1.76 billion parameters. If the claim holds, multi-task computer-aided diagnosis becomes the distribution of tiny task modules on top of one shared backbone, which changes deployment cost and opens the possibility of institutions sharing diagnostic capabilities without sharing patient data.","feed_headline":"0.17% of weights match full fine-tuning on 12 medical tasks","feed_subtitle":"UniCAD keeps a frozen 2D-trained backbone, adds tiny LoRA experts per task, and handles 2D and 3D scans together.","key_machinery":"Three mechanisms carry the argument. The Unified Embedding Layer (UEL) is a trainable patch-embedding module with two branches: one patches a 2D image $x \\in \\mathbb{R}^{H \\times W \\times C}$ with patch size $P_H \\times P_W$, the other patches a 3D volume $I \\in \\mathbb{R}^{D \\times H \\times W \\times C}$ with volumetric patches $P_D \\times P_H \\times P_W$, each branch followed by learnable positional embeddings and a CLS token; its job is to make 2D and 3D inputs conform to the token format of a frozen 2D-pretrained ViT. The second mechanism is LoRA, the low-rank adaptation identity $h_j = W_0 x_j + BA x_j$ with $B \\in \\mathbb{R}^{d \\times r}$, $A \\in \\mathbb{R}^{r \\times d}$, applied to the attention query and value projections, which confines each task's trainable parameters to a low-dimensional subspace. The third is rank standardization: since independently trained experts may use different ranks $r$, their matrices are zero-padded along the rank dimension to a common $r_{\\max}$ so all tasks in a batch share tensor shapes and the forward pass becomes $h_{ij} = W_0 x_{ij} + B'_i A'_i x_{ij}$. Around these, the proposed DEAR metric (Diagnostic Efficacy and Allocation Ratio) quantifies the system-level win as accuracy relative to a fine-tuned ViT divided by GPU memory relative to a single ViT.","core_discovery":"UniCAD's central claim is that a ViT pre-trained on natural 2D images, left entirely frozen, can serve as the shared engine of a multi-task medical diagnosis system: a trainable unified embedding layer converts both 2D images and 3D volumes into token sequences, and each task contributes only LoRA expert matrices that modify the query and value projections of every attention block via the update $\\Delta W = BA$ with $B \\in \\mathbb{R}^{d \\times r}$ and $A \\in \\mathbb{R}^{r \\times d}$. Across 11 single-label and one multi-label diagnostic tasks, including the ChestX-ray14 benchmark, these experts (0.17% of the weights) match or slightly exceed the accuracy of fully fine-tuned ViT and ResNet50 models while holding GPU memory near that of a single model: UniCAD scores DEAR 1.010 on the 11 single-label tasks versus 0.204 for individually fine-tuned ViTs. The paper further claims that a randomly ordered clinical flow of mixed 2D and 3D tasks can be served in one batched forward pass by padding token sequences to a common length and zero-padding differently ranked LoRA experts to a shared maximum rank, and that accuracy improves monotonically as the frozen backbone grows from ViT-base to ViT-giga, a scaling behavior full fine-tuning does not reliably exhibit.","pith_inferences":["If the 3D transfer premise holds, a natural testable extension is applying the same UEL recipe to non-classification tasks such as nodule segmentation, lesion detection, or report generation, where the frozen attention prior must produce spatially structured outputs rather than a single label.","The rank-standardization trick implies the ecosystem can accept experts built by different groups with different ranks, so one could push further by allowing per-layer rank allocation and measuring how much accuracy is lost to zero-padding at the batch level.","The paper never tests how the 3D positional embeddings are initialized relative to the pretrained 2D grid, so comparing inflation from 2D against random initialization would isolate how much of the 3D success belongs to the trainable embedding versus the frozen attention itself.","The DEAR exponent k=3 is introduced empirically, and the qualitative ranking of methods could shift under a different k; a sensitivity analysis would show how robust the efficiency conclusion is to that choice."],"forward_implications":["A single frozen backbone plus task modules holds GPU memory near 0.93G as the number of tasks grows, whereas loading one fine-tuned ViT per task pushes memory from 1.45G to 3.94G at eight tasks and keeps growing.","In a simulated randomly ordered clinical flow, UniCAD batches mixed 2D and 3D tasks together at batch size 32 with about 0.9 seconds of latency and 1.77G of memory, and this holds when the backbone is upgraded from ViT-base to ViT-giga.","Diagnostic accuracy improves monotonically as the frozen foundation model scales from base to giga without proportional growth in trainable parameters, so larger future vision models could be absorbed by swapping the frozen backbone alone.","Because only 0.17% of the weights are task-specific, institutions could exchange diagnostic expertise by sharing these small modules without sharing patient data or retraining the shared backbone.","On the ChestX-ray14 multi-label benchmark, UniCAD's average AUC of 0.795 edges out fully fine-tuned ViT (0.788) and ResNet50 (0.780), indicating the parameter-efficient route does not sacrifice performance on complex multi-label diagnosis."],"supporting_citations":[{"why":"Supplies LoRA, the low-rank adaptation mechanism whose 0.17% parameter budget produces each task's expert matrices.","marker":"Hu et al."},{"why":"Defines the ViT architecture that UniCAD uses as its frozen shared backbone and as the full-fine-tuning baseline it matches.","marker":"Dosovitskiy et al., 2020"},{"why":"The prior MeLo work UniCAD builds on, establishing LoRA-based multi-task diagnosis that UniCAD extends to 3D inputs.","marker":"Zhu et al., 2024"},{"why":"Supplies the MedMNIST2D and MedMNIST3D benchmarks covering 10 of the 12 evaluated diagnostic tasks.","marker":"Yang et al., 2023"},{"why":"Supplies the NIH ChestX-ray14 dataset, the multi-label benchmark where UniCAD reports the highest average AUC.","marker":"Wang et al., 2017"},{"why":"Provides the CLIP-pretrained ViT weights used to test whether UniCAD's accuracy scales with frozen backbone size.","marker":"Radford et al., 2021"},{"why":"Provides the ResNet50 architecture and ImageNet weights used as the per-task fine-tuned baseline.","marker":"He et al., 2016"}],"fun_headline_variants":["0.17% trainable weights match full fine-tuning across 12 medical tasks","Frozen ViT + 0.17% LoRA matches full fine-tuning on 12 tasks","0.17% of weights, 12 medical tasks: UniCAD matches full fine-tuning","Tiny LoRA experts on a frozen backbone: 2D and 3D CAD in one system","UniCAD: 99.83% frozen backbone, matches full fine-tuning on 12 datasets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a trainable 3D patch-embedding layer with 3D positional embeddings is enough to let the frozen, 2D-pretrained transformer blocks reason about volumetric images; if the 2D attention priors do not transfer to 3D token statistics, the reported 3D results would collapse, and the paper does not ablate this transfer or state how the 3D positional embeddings are initialized.","fun_headline_variants_meta":{"raw":{"variants":["0.17% trainable weights match full fine-tuning across 12 medical tasks","Frozen ViT + 0.17% LoRA matches full fine-tuning on 12 tasks","0.17% of weights, 12 medical tasks: UniCAD matches full fine-tuning","Tiny LoRA experts on a frozen backbone: 2D and 3D CAD in one system","UniCAD: 99.83% frozen backbone, matches full fine-tuning on 12 datasets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000805,"raw_usage":{"total_tokens":3607,"prompt_tokens":1085,"completion_tokens":2522,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":701,"completion_tokens_details":{"reasoning_tokens":2401}},"tokens_in":701,"tokens_out":2522,"duration_ms":17132,"temperature":1.0,"reasoning_tokens":2401,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:38:16.666087+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the four 3D tasks twice, initializing the UEL's 3D patch embedding and the 3D positional embeddings once by inflation or interpolation from the pretrained 2D weights and once from random values: if accuracy on NoduleMNIST3D and AdrenalMNIST3D collapses in the second case, the frozen attention's 3D transfer is load-bearing; if both runs match, the trainable embedding alone carries the result. A complementary check is a 3D dataset with strong orientation sensitivity, where a 2D attention prior might plausibly fail, which would mark the scope of the claim.","supporting_citations":[{"cited_title":", author Beyer, L","cited_arxiv_id":null,"evidence_quote":"Defines the ViT architecture that UniCAD uses as its frozen shared backbone and as the full-fine-tuning baseline it matches."},{"cited_title":", author Shen, Z","cited_arxiv_id":null,"evidence_quote":"The prior MeLo work UniCAD builds on, establishing LoRA-based multi-task diagnosis that UniCAD extends to 3D inputs."},{"cited_title":", author Shi, R","cited_arxiv_id":null,"evidence_quote":"Supplies the MedMNIST2D and MedMNIST3D benchmarks covering 10 of the 12 evaluated diagnostic tasks."},{"cited_title":", author Peng, Y","cited_arxiv_id":null,"evidence_quote":"Supplies the NIH ChestX-ray14 dataset, the multi-label benchmark where UniCAD reports the highest average AUC."},{"cited_title":", author Kim, J.W","cited_arxiv_id":null,"evidence_quote":"Provides the CLIP-pretrained ViT weights used to test whether UniCAD's accuracy scales with frozen backbone size."}],"review_version":1}