MIAP: Diverse Domains in Vision, Medicine & Astronomy

• MIAP is an acronym representing diverse research applications, including inclusive computer vision annotations, gender bias benchmarking in vision-language models, clinical outcome prediction, and radio interferometry.
• In computer vision, MIAP offers exhaustive person localization with perceived gender and age attributes that enhance fairness diagnostics and bias auditing.
• In clinical and astronomical contexts, MIAP underpins a deep learning framework with temporal attention for therapy response and a scalable multipurpose radio interferometer for astrophysical observations.

MIAP is an acronym with significant meaning in three distinct research contexts: as the “More Inclusive Annotations for People” subset in computer vision fairness research; as the “Mixed-gender Images of Annotated People” dataset used for gender-bias benchmarking in vision-LLMs; as “MIA-Prognosis,” a deep learning framework for clinical outcomes using Measurement, Intervention, and Assessment cues; and as the “Multipurpose Interferometer Array Pathfinder,” an advanced radio interferometric array. Each instantiation embodies different domain‐specific principles, methodologies, and impacts.

1. MIAP in Computer Vision: More Inclusive Annotations for People

The More Inclusive Annotations for People (MIAP) dataset is a drop-in extension of the Open Images Dataset, designed to supply exhaustive bounding-box localization and fairness-oriented perceived attribute labels for every visible person instance (Schumann et al., 2021). In contrast to the non-exhaustive, classifier-driven annotation protocol of Open Images—where only the “most specific” leaf-class-positive detections ({person, man, woman, boy, girl}) are boxed—MIAP enforces 100% bounding-box completeness for the “person” superclass in each sampled image.

Each person bounding box further receives two human-assigned attributes adopted specifically for fairness benchmarking without reliance on pre-trained attribute classifiers:

• Perceived gender presentation: {predominantly feminine, predominantly masculine, unknown}
• Perceived age range: {young, middle, older, unknown}

Labeling is constrained by stringent visual-cue-only guidelines and subject to systematic quality control with gold-standard checks and counter-stereotype examples. Box annotation policy disregards original leaf-class proposals, collapsing all subclasses into the generic “person.”

Dataset Statistics

Subset	Images	Person Boxes (orig)	Person Boxes (MIAP)	Added Boxes (%)	ρ (boxes/image)
train	70,000	—	—	33.2	—
val	7,410	—	—	24.8	—
test	22,590	—	—	23.9	—
total	100,000	357,870	454,331	27.0	orig: 3.58<br>MIAP: 4.54

Gender presentation (MIAP): 22.2% feminine, 38.3% masculine, 39.5% unknown. Age (MIAP): 6.3% young, 51.4% middle, 2.0% older, 40.3% unknown.

Fairness Diagnostics

Normalized pointwise mutual information (nPMI) analysis established that “unknown” gender, “unknown” age, and “feminine” labels exhibit the highest ΔnPMI with missing-box events, revealing systematic omission biases in the original Open Images annotation pipeline.

Research Utilization and Limitations

MIAP is intended for:

• Baseline detection and fairness comparison (original vs. exhaustive person annotations)
• Stratified evaluation by subgroup (precision, recall, AP by gender/age attribute)
• Bias auditing, including intersectional fairness Remaining caveats include non-inclusivity in attributes beyond gender/age and persistent class imbalances converging from the source data and annotation limitations (Schumann et al., 2021).

2. MIAP as a Gender Bias Benchmark in Vision-Language Evaluation

In bias evaluation literature, MIAP denotes the “Mixed-gender Images of Annotated People” subset—a curated, gender-annotated subset of OpenImages, most recently analyzed as a benchmark for feature-dependence in gender bias scoring by foundation models (Hirota et al., 9 Sep 2025). This dataset is assembled by filtering out multi-person images, retaining 5,960 single-person images (1,459 women, 4,501 men).

Spurious Correlations and Sensitivity

Quantitative analysis reveals that object co-occurrence (accuracy = 73.3%), background appearance (58.3%), and color (57.5%) are all significantly predictive of gender labels, while lighting is near chance (51.7%). Controlled perturbation experiments evaluate the stability of bias metrics (YGap for generative VLMs, MaxSkew for CLIP variants) under selective ablation of these features:

• Weak object masking (10%): up to 49% Δ in YGap
• Weak background blur: up to 88% Δ in YGap

The hierarchy of sensitivity is object > background > color > lighting, indicating that most reported “gender bias” differences in model evaluation are almost always confounded by spurious feature correlations.

Recommended Protocols

The literature proposes supplementing any bias metric on MIAP (or similar datasets) with a feature-sensitivity analysis—quantifying the average relative change Δ under perturbations. The reliability of bias measurement is thus best represented on a two-dimensional (Bias, Δ) plane, or via a composite score:

$$\beta = \text{Bias} \times (1 + \alpha \overline{\Delta})$$

where large Δ denotes unreliable bias estimation due to dataset artifact dependence (Hirota et al., 9 Sep 2025).

3. MIA-Prognosis (MIAP): Deep Learning for Clinical Therapy Response

In clinical prognosis modelling, MIAP refers to “MIA-Prognosis,” a deep learning system formalizing measurement, intervention, and assessment event-series for binary therapy response prediction (Yang et al., 2020).

Problem Formalism

For each patient, MIAP collects $M$ asynchronous time series (imaging/radiomics, blood laboratory, interventions), and a static covariate vector $s$. The therapy response label $y \in \{0,1\}$ (RECIST-based) is to be predicted at a fixed assessment time $t_*$. The prediction model $F$ is trained to minimize misclassification.

SimTA: Simple Temporal Attention

Each asynchronous sequence is encoded via the SimTA block:

• Cumulative time-interval-based attention weights:

$$A_{ij} = \begin{cases} -\lambda \sum_{k=j}^{i-1} \tau_k + \beta & \text{if } i > j \ 0 & \text{if } i = j \ -\infty & \text{if } i < j \end{cases}$$

with $\tau_i$ the inter-event intervals.

• Row-wise softmax $\alpha_{ij}$ provides causal-attention scores.
• Output is a weighted sum of ReLU-transformed feature projections, augmented with sinusoidal encodings of the recency-to-assessment $\Delta t$.

All per-modality and static embeddings are concatenated and input to a two-layer MLP. Training uses Adam with binary cross-entropy, dropout of 0.5, and feature projection size $M$0.

Empirical Results

• On synthetic asynchronous time series, SimTA achieves MSE $M$1 versus $M$2 for LSTM baselines.
• On real-world anti-PD-1 immunotherapy NSCLC patients $M$3, MIAP achieves ROC AUC = 0.80 compared to 0.70–0.71 for best LSTM. Radiomics and lab data both crucial (dropping each drops AUC to 0.47 or 0.58).
• Model stratifies survival outcomes (Kaplan–Meier, log-rank $M$4).

4. MIAP in Radio Astronomy: Multipurpose Interferometer Array Pathfinder

In the context of astronomical engineering, MIAP is the Multipurpose Interferometer Array Pathfinder, developed by the Argentine Institute of Radio Astronomy (Gonzalez et al., 2023). Phase-1 is a 16-antenna array (5 m dishes, $M$5) using radio interferometry for astronomical sources, with a current 3-antenna prototype.

System Overview

• Antenna/Drive: 5-m prime-focus dishes, alt-az mounts with independent three-phase motor control, RP2040-based control electronics, and fuzzy PI tracking. Pointing accuracy ≤0.01°.
• Receivers: Dual-polarization Vivaldi feeds, 1.2–1.45 GHz (250 MHz BW), LNAs ($M$6 K), temperature-stabilized, with calibrated noise diode injection.
• Digitization/Correlation: Each set of three antennas processed by one SNAP board (CASPER), 8-bit, 500 MS/s/channel. Downconverted/channeled with PFB/FFT (256 bins, $M$7 0.98 MHz).
• Correlation: FX architecture computes all auto and cross products, outputs to 10GbE for further calibration and imaging.

Performance and Expansion

• Maximum baseline $M$8100 m yields $M$9 arcmin beam at 1.3 GHz.
• Continuum sensitivity for 16-dish array: $s$0–5 mJy/hr.
• Scalable to $s$132–64 elements for advanced science (pulsar timing, FRBs, HI mapping).

5. Comparative Table: MIAP Instances Across Domains

Context	MIAP Expansion	Core Method/Contents	Principal Use
Computer Vision	More Inclusive Annotations for People	Exhaustive person boxes + fair attrs	Fairness-aware detection evaluation
Bias Benchmarking	Mixed-gender Images of Annotated People	Binary gender image annotations	Gender bias metric benchmarking
Clinical Modeling	MIA-Prognosis (Measurement, Intervention, Assessment)	SimTA asynchronous attention on EHR	Therapy response/outcome prediction
Astronomy	Multipurpose Interferometer Array Pathfinder	5 m radio dishes, PFB+FX correlator	Radio-astronomy, prototyping

6. Research Significance and Considerations

MIAP, in its respective instantiations, represents distinct advances: in fairness-centric corpus construction and auditing (computer vision), in robust bias benchmarking (vision-LLMs), in unified multi-modal clinical time-series prediction (biomedicine), and in reconfigurable radio telescope design (astronomy). Each use is tightly bound to domain-specific quality controls, annotation policies, or engineering architectures, and thus domain-specific best practices and limitations:

• CV MIAP: Avoids attribute classifier bias but remains limited by the scope of perceived attribute labels and source class imbalance.
• Bias MIAP: Benchmarks must be interpreted with spurious-feature sensitivity; bias metrics in isolation may be misleading.
• Clinical MIAP: SimTA provides principled handling of asynchrony/multimodality, outperforming RNNs for EHR-like data.
• Astronomy MIAP: Pathfinder validates digital back-end, noise, and tracking; scalability and modularity enable future expansion.

Each is positioned as a reference design or resource for corresponding research topics, advancing state-of-the-art approaches for their intended evaluation, monitoring, or observational use (Schumann et al., 2021, Hirota et al., 9 Sep 2025, Yang et al., 2020, Gonzalez et al., 2023).