Automating Craniofacial Reconstruction in Oral and Maxillofacial Surgery, Neurosurgery and Plastic Surgery: Dr Titipat Achakulvisut on Deep Learning Pipelines at APCA 2026

At the 14th Asian Pacific Craniofacial Association (APCA) Congress 2026, held between 28 and 30 October 2026 at Centara Grand at Central Plaza Ladprao, Bangkok, the boundaries of medical artificial intelligence took centre stage. Jointly hosted alongside the Asian Pacific Cleft Lip-Palate and Craniofacial Council (APCLPC), the Thai Cleft-Craniofacial Association (TCCA), the Society of Plastic and Reconstructive Surgeons of Thailand (ThPRS), and the Society of Aesthetic Plastic Surgeons of Thailand (ThSAPS), the congress gathered leading reconstructive surgeons and medical pioneers across the Asia-Pacific region.

Taking the stage to deliver a keynote titled "Deep Learning-Based Cranial and Mandible Reconstruction for Low-Resource Clinical Deployment", Dr Titipat Achakulvisut, Director of the Biomedical and Data Lab (BioDat Lab) at Mahidol University and AI Research Lead collaborating with Meticuly, unveiled the narrative behind a multi-year engineering journey. It is a story of how a series of mathematical breakthroughs in computer vision evolved from restoring simple cranial symmetry to automating full surgical workflows across Oral and Maxillofacial Surgery, plastic surgery and Neurosurgery.

The Clinical Challenge: Computational Bottlenecks in Virtual Surgical Planning in Oral and Maxillofacial Surgery

Modern maxillofacial and neurosurgical reconstruction has long recognised the clinical superiority of patient-matched titanium solutions over off-the-shelf plates. Yet, the traditional pathway to achieving them has harboured an unavoidable bottleneck: conventional Virtual surgical planning (VSP).

Historically, turning a patient’s raw computed tomography (CT) scan into a validated surgical guide or custom implant required hours of manual Computer-Aided Design (CAD) by specialised clinical engineers, expensive proprietary software licenses, and high-powered computational hardware. In high-volume trauma centres, district general hospitals, and resource-constrained healthcare systems, these logistical hurdles have meant that custom surgical planning remained an out-of-reach luxury for many patients.

Recognising that true clinical innovation must be accessible, Dr Titipat set out with his research team to solve a fundamental dilemma: could artificial intelligence shoulder the cognitive and geometric load of surgical design, running fast and lean enough to deploy on everyday hospital computers?

Chapter One: Landmark-Guided Defect Prediction with CraNeXt for Cranial Reconstruction

The research journey began at the top of the skull. While cranial reconstruction (cranioplasty) might appear geometrically predictable, pathological defects caused by decompressive craniectomies or high-energy trauma often present with severe, non-uniform loss of bilateral symmetry. Conventional machine learning algorithms routinely failed to extrapolate where bone ended and empty space began, or else demanded unsustainable amounts of processing memory to calculate missing geometry.

The team achieved their first major breakthrough by developing CraNeXt, published in IEEE Access. Instead of forcing a standard network to process the skull blindly, the researchers introduced an intelligent anatomical categorisation technique. By teaching the deep neural model to first identify and register surrounding cranial landmarks, the system could predict missing skull sections with remarkable structural fidelity. Infusing modern ConvNeXt blocks into a UNet architecture provided the broad perceptual field of Vision Transformers while preserving the computational efficiency of standard convolutions, enabling rapid, accurate cranial implant generation on accessible hardware.

Chapter Two: Descending to the Mandible with InceptMan

Mastering the static, dome-like curvature of the skull, however, was only the opening chapter. The real litmus test of craniofacial reconstruction lay in the lower face: the human jaw.

Unlike the cranium, the mandible is a dynamic, load-bearing bone featuring complex non-convex surfaces, delicate condylar joints, and dental occlusions. If an implant deviates by even a single millimetre, a patient’s bite alignment, mastication, and airway dynamics are compromised. Building a deep neural network capable of reconstructing missing jaw segments directly from defective CT scans without massive computing clusters was widely regarded as an elusive challenge.

The answer came with the development of InceptMan, detailed in their follow-up milestone published in IEEE Access. Drawing inspiration from InceptionNeXt architectures, the model breaks down the mandible's multi-scale structural features across parallel branches. This architecture allowed the AI to capture both the broad contour of the jawline and the intricate biomechanics of the condyles simultaneously. For the first time, deep learning could infer missing mandibular geometry end-to-end, slashing computational load while delivering the anatomical precision demanded in reconstructive Oral and Maxillofacial Surgery.

Chapter Three: The Full Leap into Automated Surgical Planning

Yet, predicting an anatomically correct bone model solved only half of the surgeon's problem. In theatre, reconstructive surgeons do not simply place an implant into a void; they must resect diseased tissue, reposition skeletal segments, and navigate osteotomy cut lines.

This realisation propelled Dr Titipat’s team toward the ultimate phase of their research: building a fully automated, clinical-grade Virtual surgical planning pipeline. Published in IEEE Access, this work integrated the entire reconstructive chain using real-world clinical datasets. Rather than functioning as an isolated design tool, the pipeline autonomously performs multi-class CT segmentation, restores missing bone geometry, simulates virtual osteotomies, and identifies optimal fixation paths.

What previously required days of iterative correspondence between surgeon and CAD technician was compressed into an automated, reproducible workflow completed in minutes. Whether applied to complex oncological resections or bimaxillary repositioning in Orthognathic Surgery, the pipeline provides surgical teams with immediate, dependable roadmaps before setting foot in the operating room.

Bringing High-Precision Care to Every Hospital

As Dr Titipat concluded his APCA presentation, the core takeaway resonated clearly with the international audience: algorithmic elegance is meaningful only when it touches patient lives.

By designing neural architectures that run efficiently on commodity clinical workstations, this research bridges the gap between deep-tech research and bedside care. The convergence of lightweight deep learning models with responsive 3D printing is breaking down economic and infrastructural barriers—ensuring that next-generation precision in Oral and Maxillofacial Surgery and Orthognathic Surgery becomes a global standard of care, rather than a regional exception.