SurvAMINN: Survival Analysis in CRLM

• SurvAMINN is a deep learning framework that utilizes autoencoder-based dimensionality reduction and multiple instance learning to convert tumor radiomics into patient-level hazard scores.
• Its methodology jointly optimizes reconstruction and survival prediction via Cox proportional hazards loss on right-censored data, addressing limitations of previous models.
• Empirical results show a C-index of 0.691, outperforming traditional radiomics approaches with pooling methods by effectively highlighting high-risk metastases.

SurvAMINN is a survival-oriented extension of AMINN introduced for postoperative prognosis in colorectal liver metastasis (CRLM). It is described as “a novel autoencoder-based multiple instance neural network for survival analysis” and as an “autoencoder-based multiple instance neural network for time-to-event survival prediction,” converting per-tumor radiomic features extracted from segmented pre- and post-contrast MRI into a patient-level hazard score for overall survival after surgery while jointly learning dimensionality reduction and survival prediction from right-censored data and emphasizing high-risk metastases (Alberb et al., 10 Sep 2025, Alberb et al., 10 Mar 2026).

1. Clinical problem and multiple-instance formulation

SurvAMINN is designed for preoperative prognosis in CRLM patients undergoing liver resection or curative-intent hepatectomy. The endpoint is postoperative overall survival, modeled as time-to-event data rather than as fixed-horizon classification. In the reported cohort, the dataset comprised 227 patients and 531 segmented tumors; the number of tumors per patient ranged from 1 to 12, the mean was 2.4 tumors per patient, and 130 patients had multifocal CRLM with mean 3.4 tumors per patient. Only 82 patients had known postoperative survival times, while the remaining patients were right-censored. This makes censored survival modeling central rather than optional (Alberb et al., 10 Sep 2025, Alberb et al., 10 Mar 2026).

The authors frame this as a multiple-instance learning (MIL) problem. A patient is the bag, tumors are the instances, and the survival outcome is patient-level. The motivation is explicitly biological and clinical: prognosis in multifocal CRLM is often driven by the most aggressive tumors rather than by the average lesion or only the largest lesion. SurvAMINN is therefore presented as a response to two limitations of prior AMINN: AMINN used 3-year survival classification rather than censored survival modeling, and it used average pooling over tumors, which can weaken the contribution of the most lethal lesions (Alberb et al., 10 Sep 2025).

A further terminological point is that the papers introducing SurvAMINN do not explicitly expand the acronym “SurvAMINN.” What is stated directly is that it extends AMINN, previously introduced as an Autoencoder-Based Multiple Instance Neural Network, into a censored survival setting (Alberb et al., 10 Mar 2026).

2. Imaging inputs, lesion-wise radiomics, and preprocessing

SurvAMINN does not ingest raw voxels. Its inputs are lesion-level radiomic feature vectors extracted independently from each tumor on preoperative MRI. The imaging sources are pre-contrast T1-weighted MRI and post-contrast T1-weighted MRI in the hepatobiliary phase after gadoxetic acid. The acquisition was 3D axial T1 imaging in pre-contrast and hepatobiliary phase with a 20-minute delay (Alberb et al., 10 Sep 2025).

Using PyRadiomics, the framework extracts 100 handcrafted radiomic features per tumor per MRI phase from the original image only. The feature classes are first-order statistics (18), shape-based (14), GLCM (22), GLRLM (16), GLSZM (16), and GLDM (14). Radiomics can be derived from either ground-truth manual tumor segmentations or predicted segmentations from the automated segmentation pipeline. Clinical variables and genomic biomarkers are compared separately in Cox analyses, but they are not described as part of the SurvAMINN input vector in the reported architecture (Alberb et al., 10 Sep 2025).

When predicted segmentations are used, very small predicted lesions are filtered out before radiomics extraction by excluding predictions with longest diameter less than or equal to the 1st percentile of training-set ground-truth tumor diameters. Before feature extraction, volumes are resampled to 2 mm isotropic spacing, intensities outside $[\mu-3\sigma,\mu+3\sigma]$ within the ROI are excluded, intensities are z-score normalized, scaled by 100, and discretized with fixed bin width 5. The extracted radiomic features are then normalized in two steps: first a shift to positive range with log transform, then z-score normalization using training-set statistics only. The reported normalization is

$$\tilde{\mathbf{x}}_i=\log\left(\max\left(\mathbf{x}_i-\min \mathbf{x}_i+\operatorname{median}(\mathbf{x}_i-\min \mathbf{x}_i),\varepsilon\right)\right),$$

followed by

$$\hat{\mathbf{x}}_i=\frac{\tilde{\mathbf{x}}_i-\boldsymbol{\mu}_i}{\boldsymbol{\sigma}_i},$$

with $\varepsilon=10^{-6}$ (Alberb et al., 10 Sep 2025).

3. Architecture and mathematical formulation

SurvAMINN is described as a fully connected neural network comprising an autoencoder for dimensionality reduction and a MIL regressor, or hazard regressor, for survival prediction from multifocal CRLM. For each lesion, the encoder maps the normalized radiomic vector to a bottleneck latent representation, the decoder reconstructs the original feature vector from that latent code, and a second branch predicts a tumor-level hazard score from the same latent representation. Exact layer widths, bottleneck size, activations, and batch size are not reported in the provided manuscript text (Alberb et al., 10 Sep 2025, Alberb et al., 10 Mar 2026).

The model’s distinctive aggregation mechanism is LogSumExp (LSE) pooling. If a patient has lesion-level hazards $\{\eta_1,\ldots,\eta_n\}$, the patient-level hazard is

$$\eta_p=\log\left(\sum_{i=1}^{n}\exp(\eta_i)\right).$$

This choice is motivated as a differentiable soft-max surrogate: if one lesion is much worse than the others it behaves like max pooling, while if several tumors are high risk they all contribute. The papers explicitly state that LSE was chosen to balance robustness and training stability while accounting for patients with multiple high-risk tumors. No attention mechanism is reported; the focus on aggressive tumors arises from lesion-level hazard prediction followed by LSE pooling (Alberb et al., 10 Sep 2025).

The autoencoder branch is trained with mean squared reconstruction error, written as

$$\mathcal{L}_{\text{MSE}}=\frac{1}{N}\sum_{i=1}^{N}|x_i-\hat{x}_i|^2$$

in one report and equivalently as lesion-wise reconstruction loss over tumors in the later report. Survival supervision is provided by Cox proportional hazards loss,

$$\mathcal{L}_{\text{CoxPH}}=- \sum_{i:\delta_i = 1} \left( \eta_i - \log \sum_{j \in \mathcal{R}(T_i)} e^{\eta_j} \right),$$

where $\delta_i$ is the event indicator, $T_i$ is the observed time, and $$\tilde{\mathbf{x}}_i=\log\left(\max\left(\mathbf{x}_i-\min \mathbf{x}_i+\operatorname{median}(\mathbf{x}_i-\min \mathbf{x}_i),\varepsilon\right)\right),$$0 is the risk set. The total objective is a weighted sum,

$$\tilde{\mathbf{x}}_i=\log\left(\max\left(\mathbf{x}_i-\min \mathbf{x}_i+\operatorname{median}(\mathbf{x}_i-\min \mathbf{x}_i),\varepsilon\right)\right),$$1

so training gradually shifts from reconstruction to survival prediction, which the authors state improves training stability (Alberb et al., 10 Sep 2025, Alberb et al., 10 Mar 2026).

At inference time, SurvAMINN outputs tumor-level hazard scores and a pooled patient-level hazard score. It does not directly output calibrated survival probabilities or a full survival curve. The patient hazard is instead used for ranking, C-index computation, Kaplan–Meier risk stratification, and univariate or multivariate Cox analyses (Alberb et al., 10 Sep 2025).

4. Role in the MRI-to-prognosis pipeline and training procedure

SurvAMINN is the final predictive component in a two-stage automated framework. The upstream segmentation pipeline segments liver, spleen, and tumors from pre- and post-contrast MRI, using partial-label learning and SAMONAI-assisted label completion; lesion-wise radiomics are then extracted from the tumor masks and passed to SurvAMINN, which outputs patient hazard. SurvAMINN depends on segmentation only indirectly, through the radiomic features computed from the masks, rather than by consuming masks or prompts directly (Alberb et al., 10 Sep 2025, Alberb et al., 10 Mar 2026).

Because pre- and post-contrast tumor segmentations were unpaired, the reported implementation trains independent pre-contrast and post-contrast SurvAMINN models and combines them by late fusion at the hazard level, or more generally by aggregating hazards from independent models. The exact algebra of this late fusion is not specified. The optimization settings given for outcome prediction are AdamW, learning rate $$\tilde{\mathbf{x}}_i=\log\left(\max\left(\mathbf{x}_i-\min \mathbf{x}_i+\operatorname{median}(\mathbf{x}_i-\min \mathbf{x}_i),\varepsilon\right)\right),$$2, weight decay $$\tilde{\mathbf{x}}_i=\log\left(\max\left(\mathbf{x}_i-\min \mathbf{x}_i+\operatorname{median}(\mathbf{x}_i-\min \mathbf{x}_i),\varepsilon\right)\right),$$3, 250 epochs, and dropout 20%; the paper also states that no GPU was required. To mitigate imbalance, each epoch is balanced by sampling an equal number of censored and uncensored patients. Survival experiments were repeated 15 times with different 3-fold cross-validation splits (Alberb et al., 10 Sep 2025).

A practical deployment issue arises when the segmentation system predicts no tumor for a patient. In that case no radiomics can be extracted and SurvAMINN cannot compute hazard. The paper notes that five patients with empty predicted segmentations were excluded from Kaplan–Meier analysis using predicted masks. This is not an architectural property of SurvAMINN itself, but it materially affects automated end-to-end use (Alberb et al., 10 Sep 2025).

5. Empirical performance and comparative evidence

The headline SurvAMINN result is a survival prediction C-index of 0.691. In the main reported survival analysis, this was accompanied by hazard ratio $$\tilde{\mathbf{x}}_i=\log\left(\max\left(\mathbf{x}_i-\min \mathbf{x}_i+\operatorname{median}(\mathbf{x}_i-\min \mathbf{x}_i),\varepsilon\right)\right),$$4, 95% confidence interval $$\tilde{\mathbf{x}}_i=\log\left(\max\left(\mathbf{x}_i-\min \mathbf{x}_i+\operatorname{median}(\mathbf{x}_i-\min \mathbf{x}_i),\varepsilon\right)\right),$$5, and $$\tilde{\mathbf{x}}_i=\log\left(\max\left(\mathbf{x}_i-\min \mathbf{x}_i+\operatorname{median}(\mathbf{x}_i-\min \mathbf{x}_i),\varepsilon\right)\right),$$6. A later report of the same automated framework summarizes the corresponding radiomics-stage result as a survival prediction C-index of 0.69 (Alberb et al., 10 Sep 2025, Alberb et al., 10 Mar 2026).

A central ablation isolates the pooling rule:

Pooling	C-index	HR (95% CI)
mean	0.556	1.114 (1.040–1.216)
largest	0.594	1.263 (1.147–1.391)
max	0.623	1.568 (1.441–1.784)
LSE	0.691	1.929 (1.890–1.980)

Only the LSE result is reported with $$\tilde{\mathbf{x}}_i=\log\left(\max\left(\mathbf{x}_i-\min \mathbf{x}_i+\operatorname{median}(\mathbf{x}_i-\min \mathbf{x}_i),\varepsilon\right)\right),$$7; the corresponding $$\tilde{\mathbf{x}}_i=\log\left(\max\left(\mathbf{x}_i-\min \mathbf{x}_i+\operatorname{median}(\mathbf{x}_i-\min \mathbf{x}_i),\varepsilon\right)\right),$$8-values for mean, largest, and max pooling are 0.759, 0.618, and 0.162, respectively. This is the paper’s most direct evidence that prognosis is better captured by model-identified high-risk tumors than by the largest-lesion heuristic or by averaging all lesions (Alberb et al., 10 Sep 2025).

SurvAMINN was also compared against conventional radiomics pipelines that separate feature reduction from prediction. Reported C-indices were 0.659, 0.647, 0.667, and 0.634 for SVM with no reduction, KBest, PCA, and mRMR, and 0.632, 0.637, 0.634, and 0.631 for RSF with the same four reduction settings. SurvAMINN’s 0.691 therefore exceeded all listed SVM and RSF baselines. Against clinical and genomic biomarkers, univariate C-indices were 0.476 for sex, 0.525 for age, 0.545 for Fong score, 0.538 for TTE or TuEn as reported in the two versions, 0.583 for APC, 0.502 for TP53, 0.538 for KRAS, and 0.500 for NRAS, versus 0.691 for SurvAMINN. In multivariate Cox analysis including all biomarkers, SurvAMINN remained the strongest and only strongly significant predictor, with reported HR $$\tilde{\mathbf{x}}_i=\log\left(\max\left(\mathbf{x}_i-\min \mathbf{x}_i+\operatorname{median}(\mathbf{x}_i-\min \mathbf{x}_i),\varepsilon\right)\right),$$9 and 95% CI $$\hat{\mathbf{x}}_i=\frac{\tilde{\mathbf{x}}_i-\boldsymbol{\mu}_i}{\boldsymbol{\sigma}_i},$$0, $$\hat{\mathbf{x}}_i=\frac{\tilde{\mathbf{x}}_i-\boldsymbol{\mu}_i}{\boldsymbol{\sigma}_i},$$1. The abstract’s statement that the framework delivers a “C-index improvement exceeding 10%” over existing clinical and genomic biomarkers corresponds to an absolute gain of 0.108 over the best compared biomarker value of 0.583 (Alberb et al., 10 Sep 2025, Alberb et al., 10 Mar 2026).

Phase ablations show that post-contrast alone performs better than pre-contrast alone and that combining pre- and post-contrast gives the best performance, with the improvement reported as statistically significant. Using predicted rather than manual segmentations caused only slight C-index drops: $$\hat{\mathbf{x}}_i=\frac{\tilde{\mathbf{x}}_i-\boldsymbol{\mu}_i}{\boldsymbol{\sigma}_i},$$2 for pre-contrast, $$\hat{\mathbf{x}}_i=\frac{\tilde{\mathbf{x}}_i-\boldsymbol{\mu}_i}{\boldsymbol{\sigma}_i},$$3 for post-contrast, and $$\hat{\mathbf{x}}_i=\frac{\tilde{\mathbf{x}}_i-\boldsymbol{\mu}_i}{\boldsymbol{\sigma}_i},$$4 when both phases were combined. Additional validation included Wilcoxon rank-sum tests for phase comparisons, median-split Kaplan–Meier analysis for high- versus low-risk groups, and a randomization test with $$\hat{\mathbf{x}}_i=\frac{\tilde{\mathbf{x}}_i-\boldsymbol{\mu}_i}{\boldsymbol{\sigma}_i},$$5 shuffled-label trainings, in which the observed C-index lay far above the null distribution (Alberb et al., 10 Sep 2025).

6. Limitations, future directions, and common misidentifications

The reported limitations are substantial. The cohort is retrospective and single-institutional, the sample size is modest for neural survival modeling, and only 82 patients had known postoperative survival times. The manuscripts do not report exact encoder and decoder layer widths, bottleneck dimension, activation functions, SurvAMINN batch size, or an explicit hyperparameter search or early stopping procedure. Dependence on segmentation quality remains operationally important because empty predicted tumor sets preclude hazard prediction. The authors also note typical radiomics concerns related to acquisition protocol, preprocessing, segmentation quality, and PyRadiomics settings, as well as the simplicity of late fusion across MRI phases when lesions are unpaired (Alberb et al., 10 Sep 2025, Alberb et al., 10 Mar 2026).

The stated future directions include adding more informative modalities such as T2-weighted MRI and CT, improving pre/post lesion matching for better fusion, integrating clinical and molecular data directly into the prediction model, using liver and spleen or background features in radiomics, improving tumor detection sensitivity so that patients with empty predicted tumor sets are not lost, and extending the framework to chemotherapy response prediction in non-surgical cohorts (Alberb et al., 10 Sep 2025).

SurvAMINN is also easy to misidentify because of lexical overlap with unrelated method names. It is not a Sharpness-Aware Minimization method. The paper “Asynchronous Sharpness-Aware Minimization for Fast and Accurate Deep Learning” explicitly states that it does not mention “SurvAMINN” by name (Jo et al., 14 Mar 2025), and “Agnostic Sharpness-Aware Minimization” likewise does not mention or define SurvAMINN (Nguyen et al., 2024). SurvAMINN instead belongs to lesion-aware radiomics-based survival modeling for CRLM, where its defining elements are censored time-to-event learning, autoencoder-based latent compression, and MIL aggregation centered on high-risk metastases (Alberb et al., 10 Sep 2025, Alberb et al., 10 Mar 2026).