---
title: 'AMIF-MDS: Wireless & Pathology Methods'
url: https://www.emergentmind.com/topics/amif-mds
type: topic
---

# AMIF-MDS: Wireless & Pathology Methods

AMIF-MDS is a context-dependent acronym used in 2025 arXiv literature for technically distinct constructs in wireless systems and computational pathology. In O-RAN testing, it denotes “Mutual Information-Driven Visualization and Clustering for Core KPI Selection in O-RAN Testing,” a pipeline that combines aggregate mutual Information in frequency (AMIF), multidimensional scaling (MDS), and DBSCAN to analyze dependencies among KPI time series [2510.02696]. In atypical mitosis research, the same acronym denotes both a training strategy for cross-domain atypical mitotic figure classification—“Atypical Mitotic Figure Multi-Domain Supervision”—and the published-name correspondence of the AMi-Br resource, described as “Atypical Mitotic Figure – Mitosis Dataset” [2508.20745], [2501.04467]. A further pathology use appears as an AMIF-MDS solution for MIDOG 2025 Track 2 built around H-optimus-0, LoRA, MixUp, soft labels, hard negative mining, adaptive focal loss, metric learning, and domain adaptation [2509.02601].

## 1. Terminological scope and referents

The acronym is used for multiple non-equivalent objects. In the O-RAN literature it names an information-theoretic analysis pipeline. In digital pathology it names both a model family and a dataset correspondence.

| Referent | Context | Source |
|---|---|---|
| “Mutual Information-Driven Visualization and Clustering for Core KPI Selection in O-RAN Testing” | KPI dependency analysis in O-RAN time-series testing | [2510.02696] |
| “Atypical Mitotic Figure Multi-Domain Supervision” | Cross-domain AMF classification in MIDOG 2025 Task 2 | [2508.20745] |
| AMIF-MDS solution | Foundation-model-driven AMF/NMF classification for MIDOG 2025 Track 2 | [2509.02601] |
| “Atypical Mitotic Figure – Mitosis Dataset” / correspondence to AMi-Br | Public dataset of normal and atypical mitotic figures on human breast cancer | [2501.04467] |

This multiplicity is not merely terminological. The O-RAN usage is centered on stochastic-process dependence, spectral decomposition, and geometric embedding. The pathology usages are centered on histologic image classification under class imbalance and domain shift, or on the curation of annotated mitotic figure patches. The acronym therefore requires domain-specific disambiguation.

## 2. AMIF-MDS in O-RAN: information-theoretic construction

In O-RAN testing, AMIF-MDS is motivated by the need to identify dependencies among performance measurements that are “inherently time-series and can be modeled as realizations of random processes” [2510.02696]. The starting point is Shannon mutual information,
$$
I(X;Y)=\int p(x,y)\log\!\left[\frac{p(x,y)}{p(x)p(y)}\right]dx\,dy,
$$
and directed information,
$$
I(X^n\to Y^n)=\sum_{t=1}^n I(X^t;Y_t\mid Y^{t-1}),
$$
which captures causal flow. The source emphasizes that both quantities are difficult to estimate reliably from continuous time-series without strong model assumptions [2510.02696].

To obtain a frequency-domain proxy that generalizes coherence to non-Gaussian processes, the method defines mutual information in frequency (MIF) through the spectral representation of a zero-mean stationary process:
$$
X(t)=\int_{-\pi}^{\pi} e^{i\omega t}\,dZ_X(\omega).
$$
The MIF between processes $X$ and $Y$ at frequencies $\omega,\nu$ is
$$
\mathrm{MIF}(\omega,\nu)\coloneqq I\!\left(dZ_X(\omega);dZ_Y(\nu)\right).
$$
Aggregate mutual Information in frequency is then defined by summing over frequency pairs:
$$
\mathrm{AMIF}(X;Y)\coloneqq \sum_{\omega\in\Omega_X}\sum_{\nu\in\Omega_Y}\mathrm{MIF}(\omega,\nu).
$$
Under mild mixing conditions, this sum recovers a quantity proportional to the mutual information rate or directed information rate [2510.02696].

The method is explicitly framed as practical rather than purely asymptotic. The paper states that finite-length, non-stationary observations are handled by selecting the most informative frequency pairs and aggregating their time-domain projections. That design choice places AMIF-MDS between classical spectral dependence analysis and direct causal-information estimation.

## 3. Quantile-based AMIF estimation, MDS embedding, and KPI clustering

The estimator operates on two real-valued time-series $X$ and $Y$ of length $T$ by partitioning each into $L$ non-overlapping segments of length $N$, applying an $N$-point FFT to each segment, and constructing frequency-wise real matrices
$$
U_j=[\Re(\hat X^{(\ell)}(f_j)),\Im(\hat X^{(\ell)}(f_j))]_{\ell=1}^L,
$$
with an analogous definition for $V_k$ on $Y$ [2510.02696]. For every frequency pair $(j,k)$, a continuous MI estimator is applied to obtain
$$
M_{jk}=I(U_j;V_k),
$$
yielding an $N\times N$ matrix $M$.

The method then introduces quantile selection. With quantile parameter $q\in(0,1]$, it sorts the $N^2$ entries of $M$, keeps the top
$$
m=\lfloor q\cdot N^2\rfloor,
$$
and defines selected frequency-index sets
$$
J=\{j:\exists k\text{ s.t. }M_{jk}\text{ is in top-}m\},\qquad
K=\{k:\exists j\text{ s.t. }M_{jk}\text{ is in top-}m\}.
$$
The aggregated matrices $X_{\mathrm sig}$ and $Y_{\mathrm sig}$ are formed by concatenating the selected $U_j$ and $V_k$, and the final AMIF estimate is
$$
\hat I=I(X_{\mathrm sig};Y_{\mathrm sig}).
$$
A normalized similarity score is then defined by
$$
S(X,Y)=\frac{\hat I}{\min\{H(X_{\mathrm sig}),H(Y_{\mathrm sig})\}},
$$
where $H(\cdot)$ is estimated marginal entropy [2510.02696].

Given $p$ time-series, pairwise similarities form a matrix $S_{ij}=S(X^{(i)},X^{(j)})$, which is symmetrized as $S=(S+S^\top)/2$ and assigned diagonal entries $S_{ii}=\infty$ because “mutual information of a continuous variable with itself is unbounded” [2510.02696]. After normalization by $s_{\max}=\max_{i<j}S_{ij}$, the method transforms similarity into dissimilarity using either a membership transform,
$$
G_{ij}=1-S'_{ij},
$$
or a logarithmic transform,
$$
G_{ij}=-\log(S'_{ij}+\epsilon),
$$
with $\epsilon\ll 1$. Classical MDS then uses
$$
B=-\tfrac12 JG^{(2)}J,\qquad J=I-\tfrac1p 11^\top,
$$
followed by the eigen-decomposition $B=V\Lambda V^\top$ and embedding
$$
Y=V_k\Lambda_k^{1/2}.
$$
DBSCAN is finally applied to the $p\times k$ embedding coordinates.

The paper reports computational complexity dominated by the MI-matrix stage, $O(p^2\cdot N^2\cdot C_{\mathrm MI})$, with $C_{\mathrm MI}}\sim O(L\log L)$ per k-NN MI estimate, followed by $O(N^2\log N^2)$ for quantile selection per pair, $O(p^3)$ for MDS, and $O(p\log p)$ for DBSCAN [2510.02696]. It also states assumptions of segmented stationarity within each FFT block, sufficient segment length $L$, and consistent k-NN parameters across pairs, while identifying limitations: only pairwise dependencies are captured, the quantile parameter $q$ trades off sensitivity against computational cost, and frequency resolution is limited by FFT length $N$.

The O-RAN case study applies the pipeline to $p=13$ PHY/MAC KPIs sampled at 20 ms intervals under random OFDM burst interference, with $N=16$, $q=1$, dissimilarity $G_{ij}=1-S'_{ij}$, embedding dimension $k=3$, and DBSCAN parameters $\epsilon=0.15$ and $\mathrm{minPts}=1$ [2510.02696]. The largest cluster contains five KPIs—$\{$MAC-DL-CQI, DL-SINR, RSRP, RSRQ, PHY-MCS$\}$—and is interpreted as the downlink link-adaptation chain. A second cluster groups $\{$RSSI, MAC-DL-PMI$\}$, while singleton clusters include SE, MAC-DL-RI, MAC-UL-Buffer, MAC-N-PRB, and DL-BLER. The reported conclusion is that the method “organically reveals the link-adaptation indicators among other clusters, and yields a ‘core’ performance measure set for future learning-driven O-RAN testing” [2510.02696].

## 4. Foundation-model-driven AMIF-MDS for atypical mitosis classification

In digital pathology, one AMIF-MDS solution addresses MIDOG 2025 Challenge Track 2 as a binary classification problem between normal mitotic figures (NMFs, label 0) and atypical mitotic figures (AMFs, label 1) [2509.02601]. The evaluation metric is Balanced Accuracy (BA), with F1-score, AUC, and per-class recall also reported. The source identifies severe class imbalance, approximately $1$ AMF:$5.74$ NMF, high intra-class morphological variability, subtle inter-class differences, and substantial domain shift across tumor types, species, scanners, and institutions [2509.02601].

After deduplication, the pooled dataset comprises MIDOG++, AMi-Br, and LUNG-MITO, totaling 11 939 crops of size approximately $64\times 64$ px from 503 whole-slide images across 10 domains. The class counts are 10 168 NMFs (85.2%) and 1 771 AMFs (14.8%) [2509.02601]. Leave-One-Domain-Out (LODO) is used for backbone/model selection and ablations, while the final run uses pooled 95% training and 5% “monitor” data for early stopping.

The backbone is H-optimus-0, described as a pathology-tuned self-supervised ViT/CNN hybrid with approximately 86 M parameters [2509.02601]. The choice is attributed to recent cross-domain generalization benchmarks and empirical LODO testing. Parameter-efficient fine-tuning is implemented through LoRA modules inserted into all transformer/CNN attention weights, with rank $r=8$, $\alpha=32$, and dropout $0.1$. Only LoRA parameters plus the final classifier and domain heads are trained, amounting to approximately 1–2% of total weights. The label head is an MLP producing a single logit $z_\theta(x)\in\mathbb R$ followed by a sigmoid, while the auxiliary domain head is a linear layer with $|D|=10$ outputs preceded by a Gradient-Reversal-Layer [2509.02601].

The training procedure combines standard MixUp with $\lambda\sim\mathrm{Beta}(0.2,0.2)$, soft labels from three-pathologist consensus, hard negative mining, adaptive focal loss, supervised contrastive metric learning, domain-aware sampling, and adversarial domain loss [2509.02601]. Standard MixUp is retained as the final choice because it yields $\Delta\mathrm{BA}\,+0.020\pm 0.017$ relative to baseline and improves AMF recall by $+0.076$. Soft labels are the average of three binary votes, encoding inter-annotator uncertainty. Hard negative mining computes
$$
d_i=\left|\sigma(z_\theta(x_i))-y_i\right|
$$
at the end of each epoch, marks the top 30% of samples by difficulty as “hard,” doubles their sampling weight in the next epoch, and refreshes the set every epoch.

The adaptive focal loss is defined for logit $z$ and soft label $y\in[0,1]$ by first setting $p=\sigma(z)$ and $p_t=y\cdot p+(1-y)\cdot(1-p)$, then using dynamic positive weight $w_+=N_{\mathrm{NMF}}/N_{\mathrm{AMF}}$ and
$$
L_{\mathrm{focal}}(z,y)=(1-p_t)^\gamma\cdot \mathrm{BCE\text{-}with\text{-}logits}(z,y;\mathrm{pos\_weight}=w_+),
$$
with $\gamma=2.0$ [2509.02601]. Embeddings $h=f_{\mathrm emb}(x)$ are $\ell_2$-normalized and optimized by a supervised contrastive loss with Multi-Similarity miner $\epsilon=0.1$ and temperature $\tau=0.1$,
$$
L_{\mathrm{con}}=-\frac{1}{|P(i)|}\sum_{p\in P(i)}\log\frac{\exp(h_i\cdot h_p/\tau)}{\sum_{a\in A(i)}\exp(h_i\cdot h_a/\tau)},
$$
combined as
$$
L_{\mathrm{total}}=L_{\mathrm{focal}}+\lambda_{\mathrm{con}}L_{\mathrm{con}},\qquad \lambda_{\mathrm{con}}=0.5.
$$

Domain-aware sampling assigns each sample a weight proportional to inverse class frequency, inverse domain frequency, and a factor of two for hard examples. An adversarial domain classification loss is applied with $\lambda_{\mathrm dom}=0.1$ and is reported to improve cross-domain BA by approximately $+0.01$ in LODO [2509.02601]. Across 10 LODO folds, the average BA is $0.851\pm 0.037$, with AMF recall $0.841\pm 0.088$ and NMF recall $0.872\pm 0.038$. The best domain is canine soft tissue sarcoma, Vienna, with BA $=0.904$; the worst is canine lymphoma with BA $=0.787$, AMF recall $0.634$, and NMF recall $0.940$. Human domains yield BA in $[0.837,0.874]$, and scanner variation is illustrated by human breast on Hamamatsu XR versus S360 with BA $0.860$ versus $0.837$ [2509.02601].

The source identifies as strengths a cross-domain BA of about $0.85$ despite severe imbalance and domain shift, stable AMF recall gains from conservative augmentations plus MixUp, and a sharper minority-class decision boundary from metric learning plus hard negative mining [2509.02601]. It also lists limitations: reduced AMF recall on the largest and most challenging domains, scanner-specific staining variations that can induce spurious domain gaps, convergence slowdown from soft-label uncertainty, and only marginal benefit from domain adversarial loss.

## 5. AMIF-MDS as Atypical Mitotic Figure Multi-Domain Supervision

A separate pathology usage expands AMIF-MDS as “Atypical Mitotic Figure Multi-Domain Supervision” and presents a training-time recipe for domain-robust AMF classification in MIDOG 2025 Task 2 [2508.20745]. The method uses a DenseNet-121 backbone without its original classification head, retaining layers up to the last convolutional block. MixStyle is inserted immediately after the first dense block and again after the third dense block. For mini-batch feature maps $x\in\mathbb R^{B\times C\times H\times W}$, sample pairs $(x,x')$ are mixed with
$$
\lambda\sim \mathrm{Beta}(\alpha,\alpha),\qquad \alpha=0.1,
$$
using channel-wise means and standard deviations and the re-standardization rule
$$
\tilde x_{c,h,w}
=\tilde \sigma_c\frac{x_{c,h,w}-\mu_c(x)}{\sigma_c(x)}+\tilde\mu_c,
$$
with $\varepsilon=10^{-6}$ for numerical stability [2508.20745]. No label mixing is performed; each transformed feature map retains the original class label.

Attention refinement is implemented with CBAM. Given raw feature map $F\in\mathbb R^{B\times C\times H\times W}$, the refined map is $\hat F=\mathrm{CBAM}(F)$ [2508.20745]. Domain labels are coarse combinations of Scanner, Origin, Species, and Tumor, with the number of domains $D$ varying by split. Global average pooling defines channel descriptors
$$
f_{i,c}=\operatorname{GAP}(\hat F_{i,c,:,:}),
$$
from which per-domain means $\mu_{d,c}$, global means $\bar\mu_c$, and inter-domain variance
$$
s_c^2=\frac1D\sum_{d=1}^D(\mu_{d,c}-\bar\mu_c)^2
$$
are computed. The alignment loss is
$$
\mathcal L_{\rm align}=\frac1C\sum_{c=1}^C\log(1+s_c^2).
$$
Its weight $\lambda_{\rm align}$ is annealed from 0 to 1 following the DANN schedule [2508.20745].

Stabilization is provided by EMA teacher distillation. After each student update,
$$
\theta_T^{(k)}=m\,\theta_T^{(k-1)}+(1-m)\,\theta_S^{(k)},\qquad m=0.999.
$$
Teacher and student logits are softened with temperature $T=2.0$, and the distillation loss is
$$
\mathcal L_{\rm KD}=T^2\,\mathrm{KL}(p_T^{(T)}\|p_S^{(T)}).
$$
The base weight is $\lambda_{\rm KD}=0.5$, warmed up linearly from 0 to 0.5 over the first 10 epochs, with gradients flowing only into the student. The overall objective is
$$
\mathcal L_{\rm total}
=\mathcal L_{\rm cls}
+\lambda_{\rm align}\mathcal L_{\rm align}
+\lambda_{\rm KD}\mathcal L_{\rm KD},
$$
where $\mathcal L_{\rm cls}$ is binary-cross-entropy with logits on the student output [2508.20745].

Training uses AdamW with initial learning rate $1\times 10^{-3}$, weight decay $1\times 10^{-2}$, batch size 512 patches balanced across classes, ReduceLROnPlateau monitoring validation balanced accuracy with minimum learning rate $1\times 10^{-5}$, gradient clipping with global norm $\le 1.0$, and early stopping after validation plateau for 10 epochs based on the EMA teacher model [2508.20745]. Domain sampling is class-balanced but not explicitly domain-balanced, because MixStyle and the alignment term are intended to enforce cross-domain mixing.

At inference time, only the student encoder, CBAM, and classification head are used. MixStyle and knowledge distillation incur zero extra test-time cost, while CBAM adds approximately 3–5% GFLOPs overhead over a plain DenseNet-121 [2508.20745]. The reported organizer-run preliminary leaderboard results are balanced accuracy $0.8762$, sensitivity $0.8873$, specificity $0.8651$, and ROC AUC $0.9499$. The authors further report a staged improvement pattern: MixStyle alone raises balanced accuracy from baseline $0.7933$ to approximately $0.82$, adding the CBAM-based alignment term increases it to approximately $0.85$, and EMA-KD yields the final $0.8762$ [2508.20745].

## 6. AMIF-MDS as dataset correspondence: AMi-Br

The dataset usage appears in the resource “Histologic Dataset of Normal and Atypical Mitotic Figures on Human Breast Cancer,” whose acronym is AMi-Br and which is explicitly described as corresponding to AMIF-MDS [2501.04467]. The resource is derived from two public breast-cancer mitosis datasets, TUPAC and MIDOG 2021, and contains 3 720 mitotic figures across 223 tumor cases. The class distribution is 2 888 normal mitotic figures (77.6%) and 832 atypical mitotic figures (22.4%) [2501.04467].

Whole-slide images are histological sections scanned at $40\times$ magnification with H&E stain using six different whole-slide scanners across three pathology centers. Annotation is performed on $128\times 128$ px image patches centered on original mitosis annotations, with three pathologists—CAB, VW, and TAD—blinded to each other’s labels [2501.04467]. The task is to classify each patch as either a normal MF, defined as one of the classical mitotic phases, or an atypical MF, defined as one of the listed atypical subtypes: polar asymmetry, chromosome segregation abnormalities, or other atypical morphologies such as dispersed fragments. Final labels are assigned by majority vote requiring at least two matching opinions.

The resource stores uniformly resized $128\times 128$ px patches, coordinates, source dataset, individual expert votes, and majority labels. The repository is public and includes image patches, CSV metadata, and Jupyter notebooks for baseline experiments [2501.04467]. The baseline study evaluates DenseNet-121 and EfficientNet V2-S under Adam optimization with learning rate $10^{-3}$ for 20 epochs, using weighted cross-entropy, focal loss, or weighted sampling with replacement. Five-fold Monte Carlo cross-validation is run under both patch-level and patient-level splits.

Balanced accuracy is defined as
$$
\mathrm{BalancedAccuracy}
=\tfrac12\left(\frac{TP}{TP+FN}+\frac{TN}{TN+FP}\right).
$$
The highest patch-level balanced accuracy is $0.806$ with EfficientNet V2-S plus weighted cross-entropy, while the highest patient-level balanced accuracy is $0.713$ with EfficientNet V2-S plus weighted sampling [2501.04467]. The reported drop from patch-level to patient-level performance indicates reliance on patient-specific cues. The resource’s stated applications include benchmarking classifiers for two-stage mitosis detection followed by AMF classification, automated quantification of AMF count and AMF ratio for prognostic studies, and transfer learning to other tumor types or stains. Its stated limitations are class imbalance, subtle morphological differences, restriction to human breast cancer, and the absence of pixel-level annotation of spindle poles or chromosomes [2501.04467].

In the pathology literature, this dataset-level meaning of AMIF-MDS interacts directly with the method-level meanings. AMi-Br appears as a source in later cross-domain training pools, including the pooled training composition for the foundation-model-driven MIDOG 2025 Track 2 solution and the AMi-Br subset used in the “Mix, Align, Distil” training pool. This suggests that, within atypical mitosis research, AMIF-MDS can denote either the data substrate, a domain-robust training framework, or a specific challenge submission, depending on context.

Source: https://www.emergentmind.com/topics/amif-mds