Nilgün Şengöz

As AI systems become more prevalent in medical diagnosis, the need for transparency and interpretability grows exponentially. In my research, I focus on Explainable AI (XAI) methods that help clinicians understand how algorithms reach their decisions.

Gradient-weighted Class Activation Mapping (Grad-CAM) is one of the key techniques I use to visualize which parts of a histopathological image the model focuses on when making a diagnosis. This visual feedback is crucial for building trust between AI systems and healthcare professionals.

The challenge is not just building accurate models — it's building models that doctors can understand, trust, and ultimately use to improve patient outcomes. This is where XAI bridges the gap between algorithmic power and clinical practice.

In our recent study on paratuberculosis diagnosis, we demonstrated that Grad-CAM heatmaps closely aligned with the regions pathologists identified as diagnostically relevant, validating the model's reasoning process.

Starting a postdoctoral journey in a new country is both exciting and challenging. Nottingham has welcomed me with its beautiful green campus, friendly people, and a vibrant research community.

The University of Nottingham offers incredible resources for researchers. The labs are well-equipped, the library is extensive, and there are countless seminars and workshops to attend.

Being part of an international research community has already broadened my perspective on AI applications. The cultural exchange, the academic environment, and the opportunity to collaborate with researchers from diverse backgrounds make this journey truly rewarding.

Living in the UK as a TÜBİTAK 2219 fellow is a unique experience that I'm grateful for every day. I look forward to sharing more about my adventures in Nottingham and beyond!

Achieving high accuracy in image classification often exceeds what a single model can deliver. That's why I've been exploring hybrid approaches that combine the feature extraction power of deep learning with the classification strength of XGBoost.

Deep learning models like VGG16 and ResNet are excellent at extracting complex visual features from images. However, feeding these features into XGBoost can further enhance classification performance beyond what either method achieves alone.

In our recent work on rotten fruit detection and histopathology classification, this hybrid approach consistently outperformed standalone models, achieving accuracy rates above 99% in several benchmarks.

The key insight is that deep learning excels at representation learning while gradient boosting excels at decision boundaries. Combining both gives us the best of both worlds.

Research context: This work was conducted at the University of Nottingham under a TÜBİTAK 2219 international postdoctoral fellowship, in collaboration with the Computational Optimisation and Learning (COL) Lab, under the supervision of Prof. Dr. Ender Özcan (University of Nottingham), with contributions from co-researcher Dr. Jeremie Clos. Accepted to the PATAT 2026 International Conference on Automated Timetabling.

The Question

Large language models are routinely described as text tools — generators of prose, code, and conversation. That framing undersells what these systems actually are at a computational level. During my postdoctoral research, I have been exploring a different question: what happens when you treat an LLM not as an output device, but as a reasoning engine embedded inside a classical optimisation algorithm?

The result is a framework I call LLM-STAR (State encoding, Target reasoning, Adaptive repair, Reward-guided selection), developed in the context of university timetabling — a classic NP-hard combinatorial scheduling problem.

Why Timetabling?

University timetabling requires assigning courses to rooms and time slots such that no student or lecturer has two commitments simultaneously, room capacities are respected, and dozens of soft preferences are satisfied as fully as possible. At scale — a university with hundreds of courses and thousands of students — this becomes computationally intractable for exact methods. Classical approaches search a vast solution space using mathematically defined moves. They evaluate, but they do not reason.

The Architecture

The LLM-STAR loop has five steps: State Encoding → LLM Reasoning → Candidate Repairs → Evaluator Scoring → State Update. The critical step is "Candidate Repairs". Where a classical metaheuristic generates neighbour solutions by random perturbation, the LLM proposes semantically motivated repairs — reasoning about patterns such as "these two courses share the majority of enrolled students — scheduling them in parallel creates a structural conflict beyond the hard constraint."

This kind of soft-constraint reasoning is difficult to formalise explicitly but straightforward to elicit from a language model. Steps 1 and 4 are classical (deterministic). Steps 2 and 3 are where the LLM contributes semantic reasoning.

Findings

Two behaviours emerged. First, on high-conflict benchmark instances, the framework reached feasible solutions in fewer iterations than stand-alone genetic algorithms. Second, even rejected LLM proposals contributed positively — the diversity of suggestions reduced the likelihood of search stagnation in local optima.

There is also an explainability dimension: because the LLM's reasoning at each step can be logged and inspected, the framework produces a natural audit trail. In regulated environments, this auditability has genuine operational value.

Broader Applicability

The finding is that LLMs can serve as first-class algorithmic components in combinatorial search. The class of problems to which this applies is large: airline crew assignment, hospital operating room scheduling, military mission planning, port berth allocation, urban traffic signal optimisation. Each is a variant of the same structure — resources assigned to time slots under hard and soft constraints, at a scale that defeats exact methods.

Note for researchers: Full methodological details will accompany the formal publication. Academic correspondence on LLM-guided search, hybrid optimisation, or combinatorial scheduling is welcome.

Research context: This work has been published in El-Cezeri Journal of Science and Engineering (ECJSE), Vol. 13, No. 2, pp. 146–160, 2026. Co-authors: G. Karaman, M. S. Çeliker, N. Y. Çan. DOI: 10.31202/ecjse.1747013

Why Standard Detectors Fail

Camouflage is one of the oldest and most persistent challenges in visual perception. Its effectiveness relies on a straightforward principle: match target appearance to background statistics so thoroughly that the human visual system finds no foothold. Standard object detection architectures are optimised for their dominant training data: objects that are visually distinct from backgrounds, with clear edges, consistent texture, and reasonable contrast. Military camouflage specifically engineers against all three of these properties.

The result: naive application of standard models produces detection rates near chance level on well-camouflaged targets. The problem is not model capacity — it is misaligned inductive bias.

The Dual-Path Attention Architecture

Our approach centres on a dual-path attention architecture that re-frames the detection task. Rather than asking "is there an object matching class X?" — the standard formulation — the architecture asks: "is there a region whose statistical properties are inconsistent with the local background model?"

This re-framing is the core insight. A camouflaged target cannot match the background perfectly — there will always be statistical residuals: micro-scale shadow inconsistencies, subtle texture regularity, edge artefacts at boundary regions.

The local path captures fine-grained texture and edge information. The global path models long-range spatial relationships and background statistics. An attention gate fusion mechanism learns dynamically which path is most informative at each spatial location. This enables the model to allocate capacity where the camouflage artefacts are strongest.

Explainability

Given my broader research programme in Explainable AI, I paid particular attention to whether the model's attention maps were interpretable by domain experts. Grad-CAM visualisations showed consistent focus on boundary regions and texture discontinuities that human experts also identified as informative. This alignment between model attention and expert judgement is a necessary — though not sufficient — condition for operational trust.

Transferability

The dual-path attention approach is not domain-specific. The model learns to find statistical inconsistencies in visual data — the same problem appears in medical imaging (subtle lesions in complex tissue), industrial inspection (surface defects that blend with material texture), remote sensing (concealed installations), and infrastructure security (foreign object detection).

The dual-path attention architecture transfers to each of these problems without modification. The domain changes what "anomaly" means. The algorithm for finding anomalies remains the same.

Looking across a decade of research, I notice a pattern I did not plan but can now articulate clearly. My publications span histopathological image classification, military camouflage detection, LLM-based combinatorial optimisation, hybrid neural architectures, dimensionality reduction, and quantum computing applications. On the surface, these seem unrelated. On closer inspection, they share a common structure: each involved designing an algorithm to find structure in data that is not obvious, then explaining why that structure exists.

The Distinction That Matters

Consider two researchers. The first is a "medical imaging AI specialist." Invaluable in that context — but when the clinical application changes, their algorithmic toolkit must adapt to new domain knowledge. The second researcher asks first: what is the structure of this problem at the level of the algorithm? They recognise that detecting a tumour in histopathological tissue and detecting a concealed object in a complex visual background are, algorithmically, the same problem — and apply the same attention-based architecture to both.

The domain changes what "anomaly" means. The algorithm for finding anomalies transfers.

Evidence from My Own Research

CLAHE contrast enhancement, developed for veterinary histopathology, transfers directly to military vision — because both involve low-contrast targets in complex backgrounds. XGBoost hybrid ensembles, developed for fruit classification, transfer to cancer detection — because both involve combining structural and appearance features for multi-class problems. Grad-CAM attention visualisation, developed for clinical XAI, transfers to defense AI — because both require operator-interpretable evidence maps for high-stakes decisions.

Each transfer emerged organically — not from planning, but from recognising that two problems with different application contexts were structurally identical at the algorithmic level.

What Algorithm Architecture Means in Practice

Problem abstraction before domain engagement. Before reading the domain literature, ask: what is the formal structure of this problem? Classification with class imbalance? Detection with low signal-to-noise? Combinatorial problem with hard and soft constraints? The formal structure determines which algorithmic family is relevant. The domain determines the data characteristics and validation criteria.

Method transfer as a first hypothesis. When facing a new problem, the first hypothesis is always: has this algorithmic structure been solved elsewhere? If so, does the solution transfer, and if not, where does it break?

Generalisability as a research contribution. A paper reporting "97% accuracy on dataset X in disease Y" has made a domain contribution. A paper showing "architecture Z outperforms baseline B on low-salience detection across three domains" has made an algorithmic contribution. The second type has a longer citation half-life.

A clarification: this is not an argument against domain expertise. It is an argument that the most productive AI researchers tend to be those whose primary fluency is in the underlying algorithms — and who engage domain knowledge as a constraint and validation mechanism rather than a primary lens.

My research in Explainable AI originated in a clinical context: making deep learning diagnostics trustworthy to pathologists. That context remains important — my CRC Press book published in 2024 addresses it directly. But over several years of research, I have become increasingly convinced that the XAI problem is not primarily a healthcare problem. It is a consequence of deploying powerful AI systems in any context where human accountability is required — which, as regulatory frameworks mature, is becoming essentially everywhere.

The Regulatory Shift

The EU AI Act, which entered into force in 2024, establishes a risk-based framework imposing transparency requirements on AI systems in proportion to potential harm. High-risk applications — medical diagnosis, credit scoring, employment screening, critical infrastructure — are subject to requirements including human oversight, model behaviour documentation, and the ability to explain individual decisions.

Globally, similar frameworks are emerging. The UAE National AI Strategy 2031 explicitly prioritises trustworthy and ethical AI. NATO AI principles require meaningful human control over autonomous systems. The FDA has issued guidance on AI medical devices. The direction of travel is consistent: AI systems making consequential decisions must be explainable.

Where XAI Requirements Are Most Acute

Healthcare & Medical Devices: EU MDR and FDA guidance require AI transparency in clinical decision support. Diagnosis, triage, and treatment recommendation systems must provide auditable reasoning to clinicians and regulators.

Defence & Autonomous Systems: NATO AI principles and international humanitarian law require meaningful human control over autonomous weapons. Explainable decision trails are a legal prerequisite.

Financial Services: GDPR Article 22 and regional governance frameworks require right-to-explanation for automated credit, insurance, and fraud decisions.

Smart City & Infrastructure: AI systems managing urban resources — traffic, utilities, emergency response — require auditable reasoning for public accountability and operator override capability.

The Methodological Toolkit

Grad-CAM produces spatial heatmaps for image classification tasks — "where was the model looking?" SHAP provides feature attribution for tabular data, grounded in game-theoretic foundations. LIME offers instance-level explanation of any black-box model when global interpretability is impossible. Attention visualisation reveals what transformer-based architectures attended to at each step.

Choosing the right method depends on three factors: the model architecture, the nature of the explanation required (spatial vs. feature vs. instance), and the audience (clinician vs. regulator vs. operator). These methods are not interchangeable.

An Honest Assessment

The most widely used post-hoc explanation methods are approximations. They reveal something true about model behaviour, but they do not provide a complete or guaranteed-faithful account. There is active and legitimate debate in the research community about whether these approximations are reliable enough for high-stakes operational deployment.

This debate is pushing the field toward intrinsically interpretable architectures — models designed for transparency from the beginning of the design process. Within five years, XAI will likely be a prerequisite for deployment approval in regulated environments rather than an optional add-on.

My CRC Press book "Explainable AI in Healthcare" (2024) was written to make these methods deployable in clinical and high-stakes settings. Academic correspondence on XAI methodology, cross-sector applications, or the regulation-explainability interface is welcome.