PaperlayerView exact revision on arXiv

Source-linked AI summary

Toward AI-Assisted Poultry Coccidiosis Diagnosis: Evaluating Gemini and BiomedParse on Eimeria Microscopy Images

Ali Alsalama, Ahmed Kubba, Manar Abu Talib

arXiv:2609.31679v1cs.CVcs.AI

TL;DR

Species-level Eimeria diagnosis is important for poultry coccidiosis management, but conventional approaches are labor-intensive and expertise-dependent, and multimodal LLMs have not been systematically assessed for this veterinary task. The study tests Gemini on microscopy images for classification and reporting and explores BiomedParse for segmentation; classification is unreliable, reports need expert validation, and segmentation is partial. The authors conclude these models are not yet suitable for standalone diagnosis.

  • Problem

    Species-level Eimeria identification matters for targeted treatment, but conventional diagnosis is labor-intensive and requires specialized expertise, while multimodal LLMs have not been systematically evaluated for this task.

    Source: §I · p. 1

  • Method

    The study evaluates Gemini on 4,225 images across seven Eimeria species for classification and pathology-report generation, and explores BiomedParse for parasite segmentation.

    Source: §I · p. 1

  • Results

    Gemini classification was unreliable with a strong default bias toward E. tenella; reports were coherent and clinically plausible but unverified, and BiomedParse segmented only parts of oocysts.

    Source: B. Pathology Report Generation · p. 3, C. BiomedParse Segmentation · p. 4, §V · p. 5

  • Takeaways & Limitations

    The authors conclude multimodal LLMs may support parasite image interpretation and reporting but are not yet suitable for standalone diagnosis.

    Source: §V · p. 5

    Treatment reports require expert veterinary validation before practical use, and segmentation was partial, requiring ground-truth masks and task-specific fine-tuning.

    Source: B. Pathology Report Generation · p. 4, C. BiomedParse Segmentation · p. 4, §V · p. 5

Abstract

from arXiv · show

Coccidiosis caused by Eimeria parasites is a major economic burden in poultry production, and effective control depends on accurate species-level diagnosis. This study evaluates whether a general-purpose multimodal large language model can support such diagnosis. Google Gemini was assessed on 4,225 mi- croscopy images covering the seven fowl-infecting Eimeria species under two prompting conditions, one without candidate labels and one with a predefined class list, and was further tested for pathology-report generation, while BiomedParse was examined for parasite segmentation. Without candidate labels, the model produced broad and taxonomically inconsistent outputs. With candidate labels, overall accuracy reached only 14.9%, with a strong bias toward E. tenella at 74% and no correct classifications for E. acervulina, E. mitis and E. praecox. Generated treatment reports were coherent but unverified, and segmentation was only partial. Current multimodal models are therefore not yet reliable for standalone Eimeria diagnosis without domain-specific fine- tuning and expert validation.

I. INTRODUCTION

The introduction frames species-level Eimeria diagnosis as important for targeted treatment but difficult to scale with conventional methods. It motivates evaluating multimodal AI for image-based diagnosis and reporting, while noting that this veterinary parasitology application has not been systematically assessed.

  • Species-level identification supports targeted treatment, while conventional microscopy and molecular diagnosis are labor-intensive and require specialized expertise.Source: §I · p. 1
  • The seven chicken-infecting Eimeria species differ in pathogenicity and intestinal targets, and their oocyst morphology makes accurate classification important for management.Source: §I · p. 1
  • Multimodal LLM use in veterinary parasitology remains largely unexplored, and Eimeria species identification and pathology reporting from microscopy images had not been systematically evaluated.The passage cites drug-resistant strains and resource-constrained settings as reasons the gap matters.Source: §I · p. 1
  • The study evaluates Gemini on 4,225 images across seven species for classification with and without candidate labels and pathology-report generation, and explores BiomedParse segmentation.Source: §I · p. 1

II. LITERATURE REVIEW

Prior work covers automated Eimeria classification and broader medical and veterinary uses of multimodal LLMs. This study evaluates an off-the-shelf general-purpose model in Eimeria microscopy, extending beyond conventional classifiers to reporting and segmentation.

  • Traditional morphology- and molecular-based Eimeria diagnosis is time-consuming, expertise-dependent, and difficult to scale, motivating automated image analysis.Source: §II · p. 1
  • Multimodal LLMs have been studied in medical imaging and veterinary tasks, but veterinary reasoning still requires domain grounding and rigorous task-specific evaluation.Source: §II · p. 2
  • Specialized Eimeria classifiers reached 96.9% accuracy with ResTFG, while the present study evaluates Gemini off the shelf without task-specific fine-tuning.Earlier reported accuracies were 85.75% for hand-crafted descriptors and 90.42% for CNN-based methods.Source: §II · p. 2
  • Automated Eimeria classification has used hand-crafted descriptors, CNNs, and other deep-learning approaches, with curated training data and per-task optimization required by specialized models.Source: §II · p. 2

III. METHODOLOGY

The study assesses models using the publicly available COCCIMORPH database of 4,225 images covering seven fowl-infecting Eimeria species. The images are heterogeneous, variably sized single-oocyst micrographs, and experiments use them without image preprocessing.

A. Eimeria Fowl Dataset

  • The dataset contains 4,225 images distributed across seven fowl-infecting Eimeria species.Table I presents the image distribution across parasite classes.Source: A. Eimeria Fowl Dataset · p. 2, Table I · p. 2
  • COCCIMORPH samples are brightfield micrographs of single sporulated oocysts, with variable dimensions, capture conditions, and class balance.Images are roughly 200–490 pixels per side, with most measuring 250–350 pixels.Source: A. Eimeria Fowl Dataset · p. 2
  • Experiments use the images as provided, without resizing, color normalization, or enhancement.Source: A. Eimeria Fowl Dataset · p. 2

B. Gemini Model

Gemini 1.5 was evaluated over the full dataset through the Gemini API, using prompts that progress from open-ended parasite recognition to constrained Eimeria classification and report generation. The prompt designs separate unconstrained recognition, fine-grained discrimination, and reporting with a supplied class.

B. Gemini Model

  • Gemini 1.5 was evaluated over the full dataset through the Gemini API, with one-word responses and retries for temporary API failures.The experiments ran in Google Colab; retries used exponential backoff for up to six attempts per image.Source: B. Gemini Model · p. 2
  • The first prompt asks for an open-ended parasite-class identification without candidate labels, testing unconstrained recognition and revealing the model’s prior.Source: B. Gemini Model · p. 2
  • The second prompt restricts classification to seven Eimeria species, enabling comparable per-class metrics within the intended hypothesis space.Source: B. Gemini Model · p. 2
  • The third prompt supplies the ground-truth parasite class and requests a pathology report, separating report generation from classification.Source: B. Gemini Model · p. 2

C. BiomedParse Model

BiomedParse is a biomedical foundation model for segmentation, detection, and recognition; it was tested on Eimeria images to explore its potential for parasite segmentation.

C. BiomedParse Model

  • BiomedParse performs segmentation, detection, and recognition across nine imaging modalities.It uses joint learning and can segment objects based on textual descriptions.Source: C. BiomedParse Model · p. 2
  • The model was trained on more than 6 million image, segmentation-mask, and textual-description triples.Source: C. BiomedParse Model · p. 2
  • BiomedParse uses natural-language labels from existing datasets and can simultaneously segment and label objects in an image.The passage reports that it excels at segmenting irregularly shaped objects.Source: C. BiomedParse Model · p. 3
  • BiomedParse was tested on Eimeria images to assess its ability to segment the parasites, despite having been trained for human medical segmentation.The test was intended to provide insight into possible fine-tuning and use for chicken parasites.Source: C. BiomedParse Model · p. 3

1) No initial information given in the input prompt:

Gemini produced broad, taxonomically inconsistent outputs without candidate labels. With a constrained species list, results were uneven and strongly biased toward E. tenella rather than reliable species recognition.

1) No initial information given in the input prompt:

2) List of classes given in the input prompt:

B. Pathology Report Generation

Gemini generated coherent, clinically plausible treatment reports, but they were not validated by a veterinary expert and require expert assessment before practical use.

B. Pathology Report Generation

  • Gemini's reports paired input oocyst images and prompts with treatment recommendations, including suggested anticoccidials and dosages that broadly matched common protocols.The examples are shown for selected Eimeria parasite classes.Source: B. Pathology Report Generation · p. 3, Figure 2 · p. 4
  • The generated reports were coherent and clinically plausible, but had not been verified by a veterinary expert.Source: B. Pathology Report Generation · p. 3
  • Expert assessment of factual accuracy, dosage correctness, and safety against established formularies is a prerequisite for clinical decision support.Source: B. Pathology Report Generation · p. 4

C. BiomedParse Segmentation

BiomedParse was tested with the prompt “Eimeria Parasite” on images from different Eimeria classes. It segmented only some parasite components, with varying success, and the authors suggest fine-tuning using manually created ground-truth masks.

C. BiomedParse Segmentation

D. Computational Cost and Inference Time

Gemini’s hosted API requires no local GPU but incurs latency and per-call costs, while specialized CNN classifiers are more efficient locally. A hardware-controlled benchmark of latency, throughput, and cost remains future work.

D. Computational Cost and Inference Time

V. CONCLUSION

Gemini classification was unreliable, with a strong default bias toward E. tenella, and current prompt-based classification is insufficient for species-level diagnosis when visual differences are subtle. Reports require expert validation, while BiomedParse segmented only parts of oocysts; the paper concludes multimodal LLMs are not yet suitable for standalone diagnosis.

V. CONCLUSION

  • Classification was unreliable, with a strong default bias toward E. tenella, making current prompt-based LLM classification insufficient for species-level diagnosis when visual differences are subtle.Source: §V · p. 5
  • Generated pathology reports were coherent and clinically plausible but require expert veterinary validation; BiomedParse segmented only parts of the oocysts.Source: §V · p. 5
  • Multimodal LLMs may support parasite image interpretation and reporting but are not yet suitable for standalone diagnostic use.Source: §V · p. 5
  • Future work will focus on dataset expansion, model fine-tuning, improved segmentation annotation, and expert-based evaluation of generated reports and treatment recommendations.Source: §V · p. 5
Loading 2609.31679v1…